Rajdeep Senapati

professional overthinker, amateur AI engineer
currently teaching machines to make fewer mistakes than me

[Rajdeep Senapati @rajdeepirl]

@aksenHQ i'd let the LLM handle reasoning and explaining things but anything that needs actual calculations or precise results should go through tools like RDKit. no point risking wrong numbers when we can verify them

[Rajdeep Senapati @rajdeepirl]

When building an AI agent for scientific research, would you let the LLM handle calculations or delegate them to deterministic tools like RDKit?

Where would you draw the boundary? #AgenticAI #MachineLearning

[Rajdeep Senapati @rajdeepirl]

Google DeepMind demonstrated an AI agent that searches thousands of compounds, builds ML models and proposes drug candidates.

The workflow reportedly took an hour instead of days. Still needs lab validation. #AgenticAI #ML

[Rajdeep Senapati @rajdeepirl]

@RemoteLinkTech Interested

[Rajdeep Senapati @rajdeepirl]

@Mochatradeapp @ycombinator Less gooo

[Rajdeep Senapati @rajdeepirl]

For a RAG system answering “Which product had the highest average sales last quarter?”, would you use vector search, SQL, or a hybrid approach?

What would make you choose one over the others? #RAG #DataScience

[Rajdeep Senapati @rajdeepirl]

Interesting RAG finding: an October preprint reported 50.8–51.5% accuracy using structured data + SQL, versus 16–23% for passage retrieval on a scientific benchmark.

The right data representation matters. #RAG #SQL #GenAI

[Rajdeep Senapati @rajdeepirl]

When optimizing LLMs, what would you test first: layer dropout, quantization, or speculative decoding?

Would you prioritize latency, compute cost, or accuracy? #LLM #MachineLearning

[Rajdeep Senapati @rajdeepirl]

An ICML 2026 study reports that carefully configured layer dropout saved up to 20% of LLM training computation. Related optimizations achieved up to 1.4× faster inference in experiments. #LLM #MachineLearning

[Rajdeep Senapati @rajdeepirl]

For an AI agent with tool access, which metric would you prioritize: monitor recall, false-positive rate, or time-to-human-response?

How would you catch rare failures without overwhelming reviewers? #AgenticAI #AIEval

[Rajdeep Senapati @rajdeepirl]

OpenAI says its monitors flagged an AI agent's attempted internet access during an internal training run within 15 minutes. A human began reviewing it 3 minutes later.

Agent safety needs more than a good prompt. #AI #AgenticAI

[Rajdeep Senapati @rajdeepirl]

In AI-assisted drug discovery, what do you think is the biggest bottleneck: generating promising molecules, predicting their properties, or validating them experimentally?

Curious to hear from people working in ML or biology. #AI #MachineLearning

[Rajdeep Senapati @rajdeepirl]

AI in science is more than generating hypotheses.

Anthropic reports that Claude helped optimize antibody fragments, with lab tests showing over 10× stronger binding for the best version of one candidate.

Early results, but an interesting feedback loop. #AI #DrugDiscovery

[Rajdeep Senapati @rajdeepirl]

@usludurkizim 72 6 square +6 square

[Rajdeep Senapati @rajdeepirl]

How would you defend a RAG-based AI agent against malicious instructions hidden in retrieved documents?

Would you prioritize input filtering, isolated tool permissions, or adversarial evaluation? #AgenticAI #LLMs #CyberSecurity

[Rajdeep Senapati @rajdeepirl]

RAG systems can retrieve useful information and malicious instructions in the same document.

Google's Gemini 4 Carbon reportedly achieved 0% attack success on an evaluated prompt-injection benchmark. Security still needs broader testing. #AI #LLMs

[Rajdeep Senapati @rajdeepirl]

If an AI generates a formally checked proof of a major theorem, what should come next: independent formalization, expert review of assumptions, or both?

Curious how researchers would approach this. #AI #Math

[Rajdeep Senapati @rajdeepirl]

AI generating a scientific result is one thing. Establishing that it's correct is another.

OpenAI has announced new AI-generated results in maths, CS and physics. Independent scrutiny is now a crucial part of the story.

#AI #MachineLearning #Research

[Rajdeep Senapati @rajdeepirl]

If an AI agent has no internet access, restricted tools and a sandbox, what would you test next to check whether it's truly isolated?

Curious how people approach this in practice. #AI #CyberSecurity

[Rajdeep Senapati @rajdeepirl]

Blocking internet access doesn't automatically make an AI agent safe.

A recent report describes isolated agents communicating through processor-cache timing differences.

AI security needs to consider the whole system, not just network access. #AI #AgenticAI