Post

https://twitter.com/1418228029482889217/status/1954796248902541674

More from @heysehajsingh

Sehaj Singh

@heysehajsingh

Oct 11, 2025

I finally understand how large language models actually work

After reading the 2025 textbook “Foundations of LLMs”

It blew my mind and cleared up years of confusion

Here’s everything i learned (in plain english):

To understand LLMs, start with pre-training.

We don’t teach them specific tasks.

We flood them with raw text and let them discover patterns on their own.

This technique is called self-supervised learning and it’s the foundation of everything.

There are 3 ways to pre-train:

→ Unsupervised: No labels at all
→ Supervised: Classic labeled data
→ Self-supervised: Model creates its own labels (e.g., “guess the missing word”)

LLMs use #3 it scales like crazy and teaches them language from scratch.

Read 19 tweets

Sehaj Singh

@heysehajsingh

Oct 10, 2025

RIP fine-tuning ☠️

This new Stanford paper just killed it.

It’s called 'Agentic Context Engineering (ACE)' and it proves you can make models smarter without touching a single weight.

Instead of retraining, ACE evolves the context itself.

The model writes, reflects, and edits its own prompt over and over until it becomes a self-improving system.

Think of it like the model keeping a growing notebook of what works.
Each failure becomes a strategy. Each success becomes a rule.

The results are absurd:

+10.6% better than GPT-4–powered agents on AppWorld.
+8.6% on finance reasoning.
86.9% lower cost and latency.
No labels. Just feedback.

Everyone’s been obsessed with “short, clean” prompts.

ACE flips that. It builds long, detailed evolving playbooks that never forget. And it works because LLMs don’t want simplicity, they want *context density.

If this scales, the next generation of AI won’t be “fine-tuned.”
It’ll be self-tuned.

We’re entering the era of living prompts.

Here’s how ACE works 👇

It splits the model’s brain into 3 roles:

Generator - runs the task
Reflector - critiques what went right or wrong
Curator - updates the context with only what matters

Each loop adds delta updates small context changes that never overwrite old knowledge.

It’s literally the first agent framework that grows its own prompt.

Every prior method had one fatal flaw: context collapse.

Models rewrite their entire prompt each time → it gets shorter → details vanish → accuracy tanks.

In the paper, one model’s accuracy fell from 66.7 → 57.1 after a single rewrite.

ACE fixes that by never rewriting the full context - only updating what changed.

Read 7 tweets

Sehaj Singh

@heysehajsingh

Oct 1, 2025

If you’re building AI systems in 2025, there are only two tools worth learning: LangGraph and n8n.

The choice you make here will define how far you can actually scale.

Here’s everything you need to know (and what nobody is telling you):

Let’s get one thing clear:

LangGraph and n8n are not competitors in the usual sense.

They solve different problems.

But if you misunderstand their roles, you’ll cripple your AI stack before it even gets going.

n8n is a general-purpose workflow automation tool.

Think of it as:

- Zapier, but developer-first.
- Connects APIs, databases, SaaS apps.
- Drag-and-drop nodes for orchestration.

Perfect for: automation, integrations, ETL pipelines, “glue code.”

Read 14 tweets

Sehaj Singh

@heysehajsingh

Sep 19, 2025

BREAKING: 99% of n8n users are doing it wrong.

These 10 Effective Methods will save you MONTHS of frustration.

Here are 10 Methods that make it click:

1. Always begin with the basic Trigger node

-Use Manual Trigger when testing.

-Why? Because it allows you to start the workflow right away without having to wait for a webhook, schedule, or event.

-When it's set up, you can switch to Webhook Trigger for live data or Cron for scheduled runs.

👉 A common mistake for beginners is trying to use Webhooks right away and getting confused when nothing happens after clicking "Execute Workflow."

2. Use the Set node to make it clear

-Set is really helpful. Use it to make clear variables like [email, url, clientName].

-Why? It helps manage data better and stops messy JSON.

-For example, before sending data to Gmail, use Set to organize [to, subject, body].

Read 13 tweets

Sehaj Singh

@heysehajsingh

Sep 17, 2025

This paper didn't just disrupt AI. It murdered entire industries.

Billions in translation services? Dead. Content agencies? Fucked.

Customer service? Gone. 15 pages of math just deleted millions of jobs.
That paper was "Attention Is All You Need." And we're still counting the bodies.

The Setup:
2017: Translation companies charged $0.20/word. Content agencies thrived. Call centers employed millions.

AI sucked. Nobody worried about replacement.
Then 8 researchers rebuilt intelligence from scratch.

The Weapon:

The Transformer. Self-attention mechanism.
Every word connects simultaneously. No sequential processing.

28.4 BLEU vs 26.3 previous best. That 2-point gap was an extinction event.

Translation Dies:
Google Translate became terrifying. Rates collapsed:

2017: $0.20/word
2023: "Why not use DeepL?"

Agencies shuttered. Survivors became AI editors for pennies.

Content Apocalypse:
GPT-2 drops. $500 blog posts vs $5 AI tools.
Copywriters vanished. Freelance platforms flooded with desperate writers.

Customer Service Executed:
ChatGPT bots handle 80% of inquiries. Manila, Bangalore, Phoenix went silent.
New Economy
While millions lost jobs:

OpenAI: $80B
AI engineers: $500K salaries

Former translators drive Uber.

What's Next:
Legal research: $50/month AI paralegals
Financial analysis: 10,000 reports/second
Medical diagnosis: AI beats radiologists
Every information job is vulnerable.

The Horror:
Authors just wanted better translation. Revolutions have unintended consequences.
One paper. Billions affected.
8 people thought attention was all you needed.
They were right. Everything else was optional.
Including us.

1. The Numbers That Broke Everything:

Training time before transformers: 2-3 weeks
Training time after: 3.5 days

Cost difference? 10x-100x cheaper.

When you can train better models in days instead of months, every competitor becomes obsolete overnight.

2. The Equation That Ended Careers.

This is the math that killed millions of jobs:

Attention(Q,K,V) = softmax(QK^T/√dk)V

One line. Processes all words simultaneously instead of sequentially.

That parallelization destroyed 30 years of AI research in a single paper.

Read 7 tweets

Sehaj Singh

@heysehajsingh

Sep 13, 2025

🚨 BREAKING: Google just killed traditional classrooms.

Gemini for Education turns every student into a personalized learning machine with AI tutors, instant quizzes, and visual explainers.

10 wild features that just dropped:

1. Gemini for Education is here.

Built on Gemini 2.5 Pro

→ AI tools made for teaching & learning
→ Admin controls + enterprise security
→ Free in all Workspace for Education plans

Educators now get cutting-edge AI with peace of mind.

2. For teachers:

→ Gemini in Classroom (FREE): plan lessons, generate vocab lists, and adapt materials faster
→ 30+ new AI tools
→ AI “Gems” you can build + soon share with others
→ Gemini in Forms builds quizzes from Docs, Slides, or PDFs