What if you could make every photograph on the internet mean something else?

When chain-of-thought prompting became a standard technique for improving LLM performance, one of the implicit promises was safety through transparency. If the model reasons step-by-step before answering, you can check the reasoning. A backdoored model, one manipulated to produce attacker-specified outputs under specific trigger conditions, would presumably give itself away: the reasoning chain would be strange, or the final answer would obviously contradict the steps leading to it.

Contributing to a production open-source project with hundreds of active contributors, a fast-moving main branch, and a review process that is simultaneously rigorous and rapid is a different experience from publishing a paper. The feedback loop is measured in hours, not months. The correctness standard is operational: it either works in deployment or it doesn’t, rather than empirical. And the thing you’re building will immediately be used by people who have no idea you exist and would not care if they did.

My name is mangled by speech recognition systems with some regularity. “Swapnil” becomes “swamp neel,” “swap nil,” or, memorably, “one pill.” This isn’t a tragedy; it’s an inconvenience. But it points at something real: ASR systems perform differently across accents, and the disparity isn’t random. It correlates with how well-represented different accents were in training data, which in turn correlates with which communities had the resources and infrastructure to generate the labeled speech data that ASR research has historically depended on.

Mechanistic interpretability has given us a way to talk about where things happen inside a language model. The induction circuit implements in-context learning. The indirect object identification circuit routes information about subjects and objects. The modular arithmetic circuit performs addition. These aren’t metaphors; they’re specific computational subgraphs, attention heads, MLP layers, residual stream connections, identified through careful ablation and activation patching experiments.

There’s an intuition, understandable and wrong, that a compressed model should be harder to attack. The argument goes roughly like this: adversarial examples exploit the model’s excessive sensitivity to high-frequency input perturbations. A pruned or quantized model has less capacity, represents simpler functions, and surely has less room for the adversarially sensitive structure that attackers exploit.

Most attacks on NLP models work by finding a perturbation tailored to a specific input: a few word substitutions that flip a sentiment classifier on one particular review, or a paraphrase that breaks a textual entailment model on one particular sentence. These attacks are powerful but narrow. They exploit the model’s behavior on a specific input rather than something fundamental about its parameters.

Problem Statement:

In most real world use cases, there exist several relational dependencies between our datasources. These relationships need to be joined and queried at search time; since these datasources are currently normalized or split logically into tables with no redundant information. Other key search service providers like Lucene and SOLR handle this by super-fast query joins using optimized Application side joins. As we will see later, we cannot use this method for our use cases due to certain restrictions. We need to find suitable alternatives for these join queries.

A new method to summarize papers; by 5 questions! This is the First Paper, that I recently studied for my application to the DeepBayes Summer School.

View of the world and the Challenge:

Technology has become inextricably intertwined with people’s lives and businesses. By predicting everything from stock price to behaviours, learning algorithms coupled with voluminous data have permeated every industry, for revenue enhancement and/or cost-cutting. But the world is sparse, signals are like grains in heaps of chaff. To make sense of these sparse signals and derive insights, creating compressed representations of real-world data is imperative.

Purpose:

To enumerate on the basic dynamics and economics of crypto-networks and its pros and cons against legacy economic systems.