AJAY KRISHNA

I share AI insights ,news and latest trends and tools - helping you stay ahead in just 5 minutes a week | @IITGuwahati @Covcampus | @UNDP Volunteer (Climate)

[AJAY KRISHNA @jaykrishAGI]

Apple Vision Pro users can now enable spatial Personas in SharePlay-enabled apps, allowing for collaboration, gaming, and media consumption with other users in a virtual space. More details 👇🏻

[Matt Dancho (Business Science) @mdancho84]

🚨BREAKING: Google just DROPPED a masterclass on GPUs

Get it here 100% free:

[Matt Dancho (Business Science) @mdancho84]

🚨 RIP Prompt Engineering

Enter Context Engineering 2.0

It completely reframes how we think about human-AI interactions.

This is what you need to know (28 page PDF):

[Harman @itsharmanjot]

A TEAM OF AI RESEARCHERS JUST OPEN-SOURCED THE BLOOMBERG TERMINAL FOR QUANT FINANCE.

A Bloomberg Terminal costs $25,000 per year per seat. Banks pay for thousands of them.

This thing reads every quant paper, every financial blog, every SEC filing, every arXiv preprint, and turns it into a searchable knowledge base. For free.

It's called QuantMind.

It just got accepted to the NeurIPS 2025 GenAI in Finance Workshop.

Here's what it actually does:

→ Ingests arXiv quant papers, financial news, blogs, and reports automatically
→ Parses PDFs, HTML, tables, and figures into structured knowledge
→ Tags every paper by research area and topic
→ Builds a semantic knowledge graph you can query in plain English
→ Plugs into DeepResearch, RAG, and MCP for multi-hop reasoning
→ Two-stage architecture: extract once, retrieve forever

Here's the wildest part:

The financial research industry publishes around 500 new papers and reports every single day.

Hedge funds pay six-figure salaries to junior analysts whose entire job is reading them.

QuantMind reads all of it. Tags it. Embeds it. Lets you ask it questions.

154 stars. 22 forks. 173 commits. MIT license. Python.

One honest note: this is a framework, not a magic alpha machine. You still need to know what to ask. But the "I haven't read that paper yet" excuse is officially dead.

The thing Wall Street charges $25,000 a year for is sitting on GitHub. Free.

Link in the comments.

[AJAY KRISHNA @jaykrishAGI]

Guess whos the odd here ?

[AJAY KRISHNA @jaykrishAGI]

Today’s agents learn. Their scaffolds don’t.

This work makes the scaffold learn too.

* Current agent scaffolds are mostly static.
* This work makes the harness self-improving.
* The result is a system where the infrastructure compounds alongside the model.

[AJAY KRISHNA @jaykrishAGI]

The biggest AI shift isn’t better answers. It’s that software is becoming on-demand. The bottleneck is moving from writing code to deciding what to build.

[How To AI @HowToAI_]

Stanford + Meta just dropped the paper that flips everything about AI agents.

It's called "Code as Agent Harness."

Right now, we treat large language models as text generators. When they need to solve a complex problem, they rely on a "chain of thought."

But natural language is slippery. It's vague. It loses context. When an agent hallucinates in English, it just keeps talking.

So they introduced a framework that changes the entire architecture of autonomy: "Code as Agent Harness."

They stopped asking the AI to reason in words, and forced it to reason in code.

Code isn't just the final output anymore. It is the memory. It is the environment. It is the boundary.

Instead of writing a paragraph about how to solve a problem, the agent writes a script, executes it, and reads the output.

Tests become its senses. Execution logs become its memory. Sandboxes become its physics.

If an agent makes a mistake in English, it apologizes and hallucinates again.

If an agent makes a mistake in code, the compiler throws an error. The trace tells it exactly what broke. The system forces it to fix it.

This is where prompt engineering dies, and systems engineering takes over.

The paper proves that reliability doesn't come from a smarter base model. It comes from the "harness" wrapped around it:

- The model proposes.
- The harness executes.
- The environment returns feedback.
- The verifier checks.

[Founder Thoughts & Strategies @mogulinfluence]

The ultimate masterclass to uno-reverse brain rot for founders:

(definitely save this)

[AJAY KRISHNA @jaykrishAGI]

For real world rag system this might be efficacious ,

Dense retrieves by meaning.

Question
↓
Embedding
↓
Meaning Search

Sparse by keywords.

Question
↓
Keyword Match
↓
Search

Hybrid combines both.

Question
↓
Dense
+
Sparse
↓
Combined Results

Ensemble combines multiple retrievers

Question
↓
Retriever A
Retriever B
Retriever C
↓
Merge Rankings

Multi-Query generates multiple search questions to improve retrieval quality

Question
↓
Generate 5 Questions
↓
Search All
↓
Merge Results

[AJAY KRISHNA @jaykrishAGI]

HyDE → Tony imagines the villain’s hideout before searching.
Auto-Merging Retriever → S.H.I.E.L.D. reads surrounding pages of a file.
RAPTOR → Nick Fury organizes intelligence into a hierarchy from reports → regions → global threat

[AJAY KRISHNA @jaykrishAGI]

Hyde vs Raptor vs Auto merging retriever

For your future diagnostic platform:

HyDE → Better retrieval for vague symptom queries (“I’m tired and dizzy”)

Generate a fake ideal answer first, then search using that answer.

Hypothetical Document Embeddings

Auto-Merging Retriever → Keep neighbouring medical context together

Retrieve small chunks, then automatically combine related chunks

Created because chunking often breaks documents into many tiny pieces.

Imagine S.H.I.E.L.D. finds

Page 23

Of a secret file

Instead of reading 1 page :

Page 22
Page 23
Page 24

RAPTOR → Search across thousands of clinical papers while preserving both detailed findings and high-level medical knowledge

Recursive Abstractive Processing for Tree-Organized Retrieval

Developed to help LLMs understand large document collections by building hierarchical summaries

Local Information
+
Global Context

Chunks
↓
Summaries
↓
Higher Summaries
↓
Tree Structure

[AJAY KRISHNA @jaykrishAGI]

Mirostat = keeps text generation balanced between boring and chaotic.

The cat sat on the mat.
The cat sat on the mat.
The cat sat on the mat.

To

The cat jumped onto the chair and watched the rain.

Seed = the starting number that makes randomness reproducible.

seed = 123

PII = personal information that can identify an individual.

[AJAY KRISHNA @jaykrishAGI]

Beam Search → Doctor Strange explores many futures.
Beam Width → How many futures he keeps.
Early Stopping → Stop when mission is solved.
No Repeat N-Gram → Deadpool can’t repeat the same joke.
Contrastive Decoding → Tony + Bruce choose the smartest answer.
Medusa Decoding → Doctor Strange sees several future steps at once

[AJAY KRISHNA @jaykrishAGI]

Beam search vs beam width vs early stopping vs no repeat N gram

The word beam comes from searching multiple promising paths simultaneously instead of following only one path.

Model predicts:

I love pizza
I love football
I love coding

Greedy deciding keeps

I love pizza

Beem search keeps :

I love pizza
I love football
I love coding

Beam Width

Beam Width = 10

Keep 10 path result will be slower though

10 futures

Early Stopping

Stop generation once a good completed answer exists.

Instead of continuing:

The capital of France is Paris.
Paris is...
Paris is...
Paris is...
Stop immediately before penultimate sentence

No Repeat N-Gram

Without protection:

I love AI.
I love AI.
I love AI.
I love AI.

With No Repeat 3-Gram:

I love AI.
AI is transforming healthcare.

[AJAY KRISHNA @jaykrishAGI]

Avengers Tower
↓
SSID (Building Name)

Iron Man Suit
↓
GPON SN (Suit Serial Number)

Tony Stark
↓
MAC Address (Personal ID)

[AJAY KRISHNA @jaykrishAGI]

Gigabit Passive Optical Network vs MAC address

Human → Passport Number

Fiber Device → GPON Serial Number

HWTC12345678

Person → Home Address

Device → MAC Address

00:1A:2B:3C:4D:5E

Your laptop connects to Wi-Fi.

The router sees its MAC address.

SSID = Wi-Fi network name.
GPON SN = Fiber device serial number.
MAC Address = Unique network hardware address of a device.

The etymology reveal itself

A service set has identified
Gigabit passive optical network
Media accessing cantrol for that unique address in device

[Amit Shekhar @amitiitbhu]

The Full Loop in Code

[AJAY KRISHNA @jaykrishAGI]

CORS
↓
"Who is allowed to talk to me?"

CSRF
↓
"Who really requested this action?"

XSS
↓
"Can attackers run code in my users' browsers?"

[AJAY KRISHNA @jaykrishAGI]

TLS → Locked briefcase.
mTLS → Both sides show ID badges.
OAuth → Temporary Avengers visitor pass.
SAML → Nick Fury vouches for you.
RBAC → Permissions based on your Avengers role.
ABAC → Permissions based on role + mission + location + conditions.

[AJAY KRISHNA @jaykrishAGI]

AI science systems don’t fail from weak search.
They fail from a small thinking frame.

Most AI stays inside fixed assumptions.
Same variables. Same rules. Same space.

This work changes that by making everything a typed artifact —
data, outputs, failures, tools, claims.

So the system can separate:

Retrieval → known facts
Search → fixed space exploration
Discovery → changing the space itself

Novelty is no longer “surprise” or “accuracy gain.”

It is:

What cannot exist in the previous system at all.

Real progress starts when AI stops only answering inside a language…
and starts rewriting the language it thinks in.