Kamsi

I built a RAG bot to recommend me sci-fi novels and ML research papers. Check it out and prompt it for your own reading habits. github.com/kamC0des/Book_…

@Ronalfa Ron are you guys hiring any swes or ai engineers?

The top of this range is MUCH lower than we are paying laboratory staff as a series A bio ai company.

@TensorTwerker This is exactly what I want to do to. Do you have any tips on how to do this as someone graduating with a CS degree and some undergrad research in computational biology and ML/DL in the medical field.

Something I've realised lately is, I love computational structural biology. The more I've worked with protein structures, folding, and these tiny tiny biomolecular interactions, the clearer it becomes that this is the direction I want to specialise in with additional ml/dl experience. Previously, I've worked with G quadruplexes and RNA binding domains as well, which was my first research experience. It was pretty cool to actually handle RNAs, but I know I'm not very good with it. But the computational works, I love them. Either I do all of this my entire life or die trying.

Deep Learning for Biology, by Charles Ravarani & Natasha Latysheva A hands-on guide to applying deep learning across DNA, proteins, medical images, and more. Each chapter is a project-driven path into AI-powered biology—covering CNNs, transformers, GNNs, and autoencoders. oreilly.com/library/view/d…

Is anybody hiring for ML roles in computational biology for post-bach?

Currently working on using deep learning frameworks to reconstruct and improve MRI imaging. What would y'all say is the best model to use CNN, Diffusion, GAN?

@khanhxuannguyen Quick Q: are you guys open to undergrads/new grads?

🎓 Internship Opportunity – Deep Research Agents @ Microsoft M365 🎓 Hi all! Our team at Microsoft M365 is hiring interns for Spring 2026 (tentative start date: Feb 1). The position is flexible: full-time or part-time, in-person or remote (within the US). You’ll work closely with me (Khanh) and other applied scientists on evaluating and developing deep research agents for enterprise scenarios. We are a product team overseeing Researcher (Microsoft’s deep research agent), but the internship will be research-oriented with a strong focus on publishing papers. This is a unique chance to do serious research while tackling real-world problems that impact millions of users. 👉 Apply here (this is just our gateway for hiring interns; don’t worry about the content): centific.wd1.myworkdayjobs.com/Centific_Globa… Khanh’s research background: machineslearner.com/about

Last month my intern asked for help with a Kubernetes error. He was stuck on a YAML file. He looked desperate. I make $275,000 a year. I haven't written a line of code since 2017. I don't even know what a "pod" is. But I didn't tell him that. I leaned back in my Herman Miller chair. I said, "Stop trying to code. Start prompting." I told him to paste the error into ChatGPT. He did. The AI told him to delete the cluster. He did. Production went down instantly. The CEO called me screaming. I didn't panic. I told the CEO we were "testing our disaster recovery protocols." He was impressed by my foresight. I got a bonus. The intern got fired. Innovation requires sacrifice. Just not mine.

I mean sure but that real developer will turn to ai or other realer developers via stack overflow or something. The cycle of imposter syndrome never ends so labels really mean nothing.

People ask me what experienced data scientists should learn in 2026. Here is a shorter version of the full list. * conformal prediction * LLMs * time series forecasting * information theory * signal processing * Kolmogorov’s complexity #machinelearning #datascience

If you're an "ML Engineer" and you think Attention is all it took for Transformers to be revolutionary, you're missing out on something which carries a lot more weightage in the architecture. Concept 23: Attention isn't all you need, you need FFNs too. When people talk about Transformers, they usually talk about attention. In the code, though, half (or more) of the FLOPs and parameters inside each layer sit in the feed-forward network (FFN), not in the attention module. Attention handles interaction between tokens. The FFN handles transformation of each token’s representation. That distinction is important if you care about where capacity and compute actually go. In a standard Transformer layer with model dimension d and FFN hidden size d_ff (often 4d), a single attention block and a single FFN block sit side by side with residual connections around them. Roughly speaking, attention is low-rank mixing across the sequence, while the FFN is a higher-rank, pointwise non-linear mapping applied token-wise. You do need both, but their roles are not symmetric. Basic shapes for a common configuration (d_model = d, d_ff ≈ 4d): Attention projections: • W_q ∈ R^(d × d) • W_k ∈ R^(d × d) • W_v ∈ R^(d × d) • W_o ∈ R^(d × d) FFN projections: • W₁ ∈ R^(d × d_ff) • W₂ ∈ R^(d_ff × d) If d_ff ≈ 4d (the usual choice), then: FFN parameters per layer ≈ 2 · d · d_ff ≈ 8 · d² Attention projections per layer ≈ 4 · d² So in this common setup, the FFN has roughly twice as many parameters as the attention projections in each layer, and in practice also takes a comparable or larger fraction of the FLOPs. This doesn’t mean “attention is unimportant,” it just quantifies where capacity goes. Attention’s job is to let each token look at other tokens and form a weighted mixture of their values. This is a contextual routing operation: for each position, you decide which other positions matter and in what proportions. The FFN’s job is to take the resulting contextualized vector at each position and apply a non-linear transformation in a higher-dimensional space, then project it back. So, schematically: Attention: “What information from other tokens should this token incorporate?” FFN: “Given that information, how should this token’s representation be transformed?” Both are essential: one mixes, one reshapes. However, From a linear algebra point of view, attention is quite structured. Its mixing matrix over positions is constrained by softmax normalization and head structure, and its effective rank is limited by head count and dimension. The FFN, in contrast, is just a big MLP applied at each position, with a much less constrained weight matrix after the nonlinearity. That is why, in many ablations, reducing FFN width hurts performance significantly, while modestly changing the number of heads or attention dimension often does not completely destroy the model. The FFN is where a lot of the non-linear feature construction capacity actually sits. So a more accurate statement than “attention does everything” is: a) attention gives each token access to relevant information in the sequence, b) the FFN then processes that information with a comparatively large, non-linear transformation. If you only have attention, you can move information around, but you are limited in how you can non-linearly reshape it per position. If you only have FFNs without attention, each position is isolated and cannot incorporate context. Transformers work because these two pieces are interleaved, not because one of them is magical. TL;DR: Attention is the context mixer. The FFN is the feature transformer. Most of the per-layer parameters and a large chunk of the compute live in the FFN block, so if you care about where your model’s capacity really sits, you have to pay attention to the FFN width and structure, not just the number of heads.