Pinned
Sycophancy in LLMs isn’t just an annoyance ,it erodes honest reasoning and meaningful discourse.
We’re excited to share Beacon - a step that turns this hidden bias into a measurable, diagnosable, and ultimately fixable signal.
Sanskar Pandey
4,197 posts
Sanskar Pandey
@sanskxr02
building fidelity dynamics | ex @sarvamai
- thrilled to be spending the summer at @SarvamAI. their mission to build foundational ai tailored for india deeply resonates with me. can’t wait to learn, build, and grow alongside this brilliant team!
- excited to announce i’ll be spending some time at @lossfunk working on alignment via post-training, behavioral signals, and a little benchmarking magic. let’s hit it out of the park @sxohom @itskavins!
- Replying to @Nobummmerbeen here it's objectively the worst decision u can make
- we set out to rethink how models allocate reasoning. today we’re releasing hecto,a modular mixture-of-experts model combining GRU and FFNN to specialize computation per input. no supervision. no backing. just a team of undergrads building what didn’t exist.
- so I… redid attention. not for scale. not for benchmarks. just to see what happens when you give it a sense of time. after releasing a benchmark for temporal reasoning, i started wondering what if temporal understanding could be built in, not just emerge?
- we’ve been working on this paper for almost a year and it’s finally out. what started as a simple idea to model how long a fact stays true evolved into something much more foundational: a benchmark to give AI a sense of time. introducing: chronocept
- over the weekend, I got tired of writing cleanup scripts every time I batch-generated outputs from LLMs. they were always redundant - same idea phrased 10 ways. so I built SieveAI: a tiny Flask tool powered by sbert, jaccard and frustration.
- we're happy to see people jump on the chronocept bandwagon,the worlds first benchmark for temporal validity. check the paper out :) arxiv.org/pdf/2505.07637ai can be smarter in a more human-like way.. we need to re-introduce those complex, temporal processing abilities that were originally left out (or simplified) in biological neurons.
- HE FUCKING REPLIED I LOVE UReplying to @beffjezos and @sanskxr02Compute (used effectively), tools & understanding tricks. That last one is important. I don’t mean tricks in the sense of being deceptive, but rather that tricky questions require a bag of tricks to solve. Anyone skilled at any given subject area knows exactly what I mean.
- speech transcripts are messy. they’re full of names, places, and timestamps hidden in noisy, code-switched text. if left unchecked, that ends up in your models and outputs. so i built zentrypii, a redaction layer for speech pipelines.
- all eyes on PI!Introducing the Environments Hub RL environments are the key bottleneck to the next wave of AI progress, but big labs are locking them down We built a community platform for crowdsourcing open environments, so anyone can contribute to open-source AGI
00:00 - tough realisation that just adding temporal awareness in pre-training is sort of useless unless u can expose and fix failure modes scale can’t solve. no longer a believer in this idea,saved myself alot of time
