DEV Community

Devanshu Biswas profile picture

Devanshu Biswas

404 bio not found

Location Mumbai, India Joined Joined on  Email address dev48vb@gmail.com Personal website https://dev48.infy.uk github website
A Search Scope Is a Read Scope: 751 Yes/No Questions Recover Every "Private" Field in My Personal API
dev48v profile

A Search Scope Is a Read Scope: 751 Yes/No Questions Recover Every "Private" Field in My Personal API

#security #python #api #opensource
Comments
4 min read

Want to connect with Devanshu Biswas?

Create an account to connect with Devanshu Biswas. You can also sign in below to proceed if you already have an account.

Create Account
Already have an account? Sign in
Your Vector Store Rejects the Embedder Change That Barely Matters and Cannot See the One That Destroys Retrieval
dev48v profile

Your Vector Store Rejects the Embedder Change That Barely Matters and Cannot See the One That Destroys Retrieval

#ai #rag #embeddings #programming
Comments
4 min read
MC Dropout With 10 Samples Is Further From the Truth Than Just Turning Dropout Off. Here Is the Break-Even.
dev48v profile

MC Dropout With 10 Samples Is Further From the Truth Than Just Turning Dropout Off. Here Is the Break-Even.

#deeplearning #machinelearning #ai #python
Comments
4 min read
Your 90%-Precision Threshold Delivers 63.5% After a Prevalence Shift — With TPR and FPR Bit-Identical
dev48v profile

Your 90%-Precision Threshold Delivers 63.5% After a Prevalence Shift — With TPR and FPR Bit-Identical

#machinelearning #datascience #python #programming
Comments
3 min read
I Implemented npm, Cargo, Composer and pip Semver. Only 2 of 13 Ranges Mean the Same Thing in All Four.
dev48v profile

I Implemented npm, Cargo, Composer and pip Semver. Only 2 of 13 Ranges Mean the Same Thing in All Four.

#javascript #npm #rust #programming
Comments
4 min read
I Built a Rig to Show That an Agent's Own Notes Poison Its Context. The Effect Was Not There.
dev48v profile

I Built a Rig to Show That an Agent's Own Notes Poison Its Context. The Effect Was Not There.

#ai #llm #agents #programming
Comments
3 min read
The Optimistic-UI Rollback Everyone Writes Is Exactly Right for a Toggle and the Worst Option for a Set
dev48v profile

The Optimistic-UI Rollback Everyone Writes Is Exactly Right for a Toggle and the Worst Option for a Set

#webdev #javascript #ux #frontend
Comments
3 min read
Whatever Length-3 Coin Pattern You Pick, I Have a Reply That Beats It — and My Worst Case Is 2/3
dev48v profile

Whatever Length-3 Coin Pattern You Pick, I Have a Reply That Beats It — and My Worst Case Is 2/3

#gamedev #javascript #probability #math
Comments
3 min read
The Change Detector Everyone Writes First Alerts on 48 of 48 Fetches
dev48v profile

The Change Detector Everyone Writes First Alerts on 48 of 48 Fetches

#python #opensource #programming #webdev
Comments
3 min read
Retries Take Your Agent's Success Rate to 96.95% While Exactly-Once Goes Down
dev48v profile

Retries Take Your Agent's Success Rate to 96.95% While Exactly-Once Goes Down

#ai #llm #distributedsystems #programming
Comments
2 min read
Raising the Batch Size Makes Batch Norm's Inference Statistics 13x Worse
dev48v profile

Raising the Batch Size Makes Batch Norm's Inference Statistics 13x Worse

#deeplearning #machinelearning #python #ai
Comments
2 min read
Split Before You Scale Guards the Mistake That Costs 0.000168. The One That Costs 0.64 Gets a Footnote.
dev48v profile

Split Before You Scale Guards the Mistake That Costs 0.000168. The One That Costs 0.64 Gets a Footnote.

#machinelearning #datascience #python #programming
Comments
2 min read
Zero of Ten Ignore Patterns Mean the Same Thing in git, rsync, a Shell Glob and Docker
dev48v profile

Zero of Ten Ignore Patterns Mean the Same Thing in git, rsync, a Shell Glob and Docker

#git #programming #linux #devops
Comments
2 min read
Your Schema Says 65% Valid. 45% Are True. The Repair Layer Makes Both Numbers Worse.
dev48v profile

Your Schema Says 65% Valid. 45% Are True. The Repair Layer Makes Both Numbers Worse.

#ai #llm #programming #json
Comments
2 min read
Save scrollTop Is Exact on 600 of 600 Journeys, and 0 of 600 Once Three Items Arrive
dev48v profile

Save scrollTop Is Exact on 600 of 600 Journeys, and 0 of 600 Once Three Items Arrive

#webdev #javascript #frontend #ux
Comments
2 min read
In Undercut the Optimal Strategy Never Names 1 or 2, and the Closed Form Breaks at n=7
dev48v profile

In Undercut the Optimal Strategy Never Names 1 or 2, and the Closed Form Breaks at n=7

#algorithms #math #javascript #programming
Comments
2 min read
Read Your Photos' Modification Times and 1,007 of 1,513 Land in the Month You Copied the Library
dev48v profile

Read Your Photos' Modification Times and 1,007 of 1,513 Land in the Month You Copied the Library

#python #opensource #privacy #programming
Comments
2 min read
Your Vector Index Spends 1.40x More on the Queries It Already Answers Perfectly
dev48v profile

Your Vector Index Spends 1.40x More on the Queries It Already Answers Perfectly

#ai #database #performance #programming
Comments
3 min read
Rotate the Coordinates and EWC's Penalty Loses 84.8% of Its Value. The Function Never Changed.
dev48v profile

Rotate the Coordinates and EWC's Penalty Loses 84.8% of Its Value. The Function Never Changed.

#deeplearning #machinelearning #python #ai
Comments
2 min read
The Exactly-Correct Model Is First on Brier, Log Loss, AUC and Accuracy — and 5th of 10 on F1@0.5
dev48v profile

The Exactly-Correct Model Is First on Brier, Log Loss, AUC and Accuracy — and 5th of 10 on F1@0.5

#machinelearning #datascience #python #programming
Comments
2 min read
Put the Longer Branch First Is a No-Op in grep and Changes 15.5% of Answers in JavaScript
dev48v profile

Put the Longer Branch First Is a No-Op in grep and Changes 15.5% of Answers in JavaScript

#regex #programming #javascript #linux
Comments
2 min read
The Logit-Bias Ban in Every Provider's Docs Leaves the Word Writable 81% of the Time
dev48v profile

The Logit-Bias Ban in Every Provider's Docs Leaves the Word Writable 81% of the Time

#ai #llm #python #programming
Comments
2 min read
Your Focus-Trap Tests Pass for 3,600 of 3,600 Random Element Orders, Which Is Why Six of Seven Traps Are Wrong
dev48v profile

Your Focus-Trap Tests Pass for 3,600 of 3,600 Random Element Orders, Which Is Why Six of Seven Traps Are Wrong

#webdev #javascript #a11y #frontend
Comments
2 min read
The Only Rule Anybody States for Sim Is 570x Ambiguous, and Its Worst Reading Is 56x Worse Than Random
dev48v profile

The Only Rule Anybody States for Sim Is 570x Ambiguous, and Its Worst Reading Is 56x Worse Than Random

#algorithms #math #javascript #gamedev
Comments
2 min read
Your Search Says Five Sources. It Is Standing on 4.75 Documents, and Some of Them Quote Each Other
dev48v profile

Your Search Says Five Sources. It Is Standing on 4.75 Documents, and Some of Them Quote Each Other

#python #opensource #ai #programming
Comments
4 min read
Widen the Confirmation Window and Every Response-Time Metric Improves While Harm Goes Up 1.65x
dev48v profile

Widen the Confirmation Window and Every Response-Time Metric Improves While Harm Goes Up 1.65x

#sre #devops #python #opensource
Comments
2 min read
Halve the Model's Error Rate and More Wrong Answers Reach the Customer. Same Reviewer.
dev48v profile

Halve the Model's Error Rate and More Wrong Answers Reach the Customer. Same Reviewer.

#ai #llm #productivity #programming
Comments
2 min read
Averaging Predictions Is a Theorem. Averaging Weights Is a Coincidence That Usually Works.
dev48v profile

Averaging Predictions Is a Theorem. Averaging Weights Is a Coincidence That Usually Works.

#deeplearning #machinelearning #python #programming
Comments
2 min read
What ε Actually Buys: Checking a Privacy Guarantee Against Its Own Definition, Not a Simulation
dev48v profile

What ε Actually Buys: Checking a Privacy Guarantee Against Its Own Definition, Not a Simulation

#machinelearning #privacy #datascience #security
Comments
2 min read
NFC(NFD(c)) Fails for 1,120 of the 2,061 Characters That Have a Canonical Decomposition
dev48v profile

NFC(NFD(c)) Fails for 1,120 of the 2,061 Characters That Have a Canonical Decomposition

#unicode #programming #javascript #i18n
Comments
2 min read
The Same Temperature, top_p and top_k Do Not Mean the Same Thing in Two Mainstream Stacks
dev48v profile

The Same Temperature, top_p and top_k Do Not Mean the Same Thing in Two Mainstream Stacks

#ai #llm #programming #python
Comments
2 min read
The Drag Threshold in the Most Code Calls 959 of 960 Deliberate Clicks a Drag
dev48v profile

The Drag Threshold in the Most Code Calls 959 of 960 Deliberate Clicks a Drag

#webdev #javascript #frontend #ux
Comments
2 min read
In Wythoff's Game, Playing the Nim Strategy Is 22x Worse Than Closing Your Eyes
dev48v profile

In Wythoff's Game, Playing the Nim Strategy Is 22x Worse Than Closing Your Eyes

#algorithms #math #javascript #programming
Comments
2 min read
Widen the Worker Pool From 1 to 24 and Every Number the Broker Reports Is Identical. Lost Updates Go 0 to 1,710.
dev48v profile

Widen the Worker Pool From 1 to 24 and Every Number the Broker Reports Is Identical. Lost Updates Go 0 to 1,710.

#devops #ai #python #distributedsystems
Comments
2 min read
Temperature 0, a Fixed Seed, and a Different Reply Anyway — Because of Who Else Was in Your Batch
dev48v profile

Temperature 0, a Fixed Seed, and a Different Reply Anyway — Because of Who Else Was in Your Batch

#ai #llm #programming #machinelearning
Comments
1 min read
The Tornado Chart Is Blind to the Thing That Goes Wrong, Because the Base Case Is Where Nothing Does
dev48v profile

The Tornado Chart Is Blind to the Thing That Goes Wrong, Because the Base Case Is Where Nothing Does

#python #opensource #programming #datascience
Comments
2 min read
The Softmax Bottleneck Is an Integer, and the Thing It Predicts Is Continuous
dev48v profile

The Softmax Bottleneck Is an Integer, and the Thing It Predicts Is Continuous

#machinelearning #deeplearning #ai #python
Comments
2 min read
A Generalization Bound Is Correct, and This Page Measures What It Is Worth
dev48v profile

A Generalization Bound Is Correct, and This Page Measures What It Is Worth

#machinelearning #statistics #datascience #math
Comments
2 min read
CREATE TABLE p(x FLOATING POINT) Is an INTEGER Column
dev48v profile

CREATE TABLE p(x FLOATING POINT) Is an INTEGER Column

#sqlite #database #programming #javascript
Comments
1 min read
Line-Range Edits Are Wrong on 93.14% of Three-Hunk Calls and Refuse 0.00% of Them
dev48v profile

Line-Range Edits Are Wrong on 93.14% of Three-Hunk Calls and Refuse 0.00% of Them

#ai #llm #programming #python
Comments
2 min read
Five of Eight Input Masks Pass Every Test, Four of Those Five Are Wrong, and Backspace Does Nothing in Exactly Three Places
dev48v profile

Five of Eight Input Masks Pass Every Test, Four of Those Five Are Wrong, and Backspace Does Nothing in Exactly Three Places

#webdev #javascript #frontend #ux
Comments
2 min read
Counting the Edges in Hackenbush Is Exactly a Coin Flip, at Every Even Length, Forever
dev48v profile

Counting the Edges in Hackenbush Is Exactly a Coin Flip, at Every Even Length, Forever

#algorithms #math #javascript #programming
Comments
2 min read
Log Coverage Is 1.0000 at Every Schema Width, Including the Width Where the Record Is Empty
dev48v profile

Log Coverage Is 1.0000 at Every Schema Width, Including the Width Where the Record Is Empty

#ai #devops #python #observability
Comments
2 min read
Your Decontamination Report Measures the N-Gram Overlap It Happened to Look For
dev48v profile

Your Decontamination Report Measures the N-Gram Overlap It Happened to Look For

#ai #llm #machinelearning #datascience
Comments
2 min read
Four of Six Self-Checks Cannot Fail, and the Two That Can Go Red on the Better Configurations
dev48v profile

Four of Six Self-Checks Cannot Fail, and the Two That Can Go Red on the Better Configurations

#python #opensource #ai #programming
Comments
2 min read
The Honest Gradient Through a Binary Bottleneck Is Not Small. It Is Exactly Zero on 42 of 42 Coordinates.
dev48v profile

The Honest Gradient Through a Binary Bottleneck Is Not Small. It Is Exactly Zero on 42 of 42 Coordinates.

#machinelearning #deeplearning #ai #python
Comments
2 min read
Two Groups With Identical ROC Curves, Identical TPR and FPR to the Last Bit, and a 13-Point Selection Gap
dev48v profile

Two Groups With Identical ROC Curves, Identical TPR and FPR to the Last Bit, and a 13-Point Selection Gap

#machinelearning #datascience #statistics #python
Comments
2 min read
5.21% of DEFLATE Matches Copy Bytes That Have Not Been Written Yet
dev48v profile

5.21% of DEFLATE Matches Copy Bytes That Have Not Been Written Yet

#programming #algorithms #javascript #compression
Comments
2 min read
71.53% of Prompt Cut Positions Are Not a Token Boundary of the Text That Follows
dev48v profile

71.53% of Prompt Cut Positions Are Not a Token Boundary of the Text That Follows

#ai #llm #programming #python
Comments
2 min read
Seven of Nine Scroll-Spy Implementations Pass the Only Test Anyone Writes, and Six of Those Seven Are Wrong
dev48v profile

Seven of Nine Scroll-Spy Implementations Pass the Only Test Anyone Writes, and Six of Those Seven Are Wrong

#webdev #javascript #css #frontend
Comments
2 min read
A Grid Search for Kuhn Poker's Optimal Strategy Finds Either q/3 + 1 Answers or Exactly Zero
dev48v profile

A Grid Search for Kuhn Poker's Optimal Strategy Finds Either q/3 + 1 Answers or Exactly Zero

#algorithms #math #javascript #programming
Comments
2 min read
Twelve of the Twenty-Four Sandbox Settings Do More Harm Than Having No Sandbox At All
dev48v profile

Twelve of the Twenty-Four Sandbox Settings Do More Harm Than Having No Sandbox At All

#ai #security #python #devops
Comments
2 min read
Context Utilisation Separates Six Packers by 2.45%. Whether the Answer Survived Separates Them by 56.56%.
dev48v profile

Context Utilisation Separates Six Packers by 2.45%. Whether the Answer Survived Separates Them by 56.56%.

#ai #llm #javascript #programming
Comments
2 min read
The Recogniser With the 36% Worse Word Error Rate Retrieves Better
dev48v profile

The Recogniser With the 36% Worse Word Error Rate Retrieves Better

#python #opensource #ai #programming
Comments
2 min read
The LoRA Rank That Wins Is Never the Rank the Task Actually Has
dev48v profile

The LoRA Rank That Wins Is Never the Rank the Task Actually Has

#machinelearning #ai #python #deeplearning
Comments
2 min read
A Clustering That Knows Nothing Scores 0.5033. One That Gets Half the Items Right Scores 0.3118.
dev48v profile

A Clustering That Knows Nothing Scores 0.5033. One That Gets Half the Items Right Scores 0.3118.

#machinelearning #datascience #python #statistics
Comments
2 min read
The tar Path Limit Is Not 255 Bytes, and the Wrong Number Scores Better Than the Right One
dev48v profile

The tar Path Limit Is Not 255 Bytes, and the Wrong Number Scores Better Than the Right One

#programming #linux #javascript #devops
Comments
2 min read
Your Stop Sequence Fires on 100% of the Replies It Must Not Touch and 0% of the Ones It Was Installed For
dev48v profile

Your Stop Sequence Fires on 100% of the Replies It Must Not Touch and 0% of the Ones It Was Installed For

#ai #llm #programming #python
Comments
2 min read
Five of Eight Toolbar Implementations Satisfy the Only Property Anyone Tests, and Three of Those Five Are Wrong
dev48v profile

Five of Eight Toolbar Implementations Satisfy the Only Property Anyone Tests, and Three of Those Five Are Wrong

#webdev #javascript #css #frontend
Comments
2 min read
Hold at 20 Is a Theorem, and It Is Not Even a Local Maximum
dev48v profile

Hold at 20 Is a Theorem, and It Is Not Even a Local Maximum

#algorithms #math #javascript #programming
Comments
2 min read
loading...