Pranjal Sahu

Senior Research Scientist at @siemenshealth | PhD in CS @stonybrooku Medical Imaging, Computer Vision

[Pranjal Sahu @pranjalsahu]

ITK 5.3.0 has been released! https://t.co/te9uH1PqyI

[alphaXiv @askalphaxiv]

“Program-as-Weights”

LLMs are great at fuzzy functions like log triage, JSON repair, and intent classification, but calling a big model on every input is slow, expensive, and not local.

This paper compiles the fuzzy function once from a natural language spec into a small neural program, a pseudo-program plus a LoRA adapter.

A frozen 0.6B interpreter then runs it locally, matching Qwen3 32B prompting on FuzzyBench while using about 50x less inference memory and running around 30 tok/s on a MacBook M3!

[Pranjal Sahu @pranjalsahu]

MRI explained: How does it work? https://t.co/syqfMtrrLe via @YouTube

[Vik Paruchuri @VikParuchuri]

We're open sourcing a 9B model that extracts structured data from documents at near-frontier performance.

- 90.2% on our bench, vs Gemini 3.5 Flash at 91.3%
- Leads extraction models like NuExtract3 (81.5%)
- 9.5s p50 timings
- Pass JSON schema

[Rahul @sairahul1]

Andrej Karpathy: "90% of Claude's mistakes come from missing context, not a weak model."

41% mistake rate without a CLAUDE.md. 11% with the 4-rule baseline. 3% with the 12-rule version below

here are the 12 rules senior engineers settled on:

1. think before coding: state assumptions, don't guess. the model can't read your mind, stop hoping it will

2. simplicity first: minimum code, no speculative abstractions. the moment you let Claude add "for future flexibility," you've added 200 lines you'll delete next quarter

3. surgical changes: touch only what you must. don't let it improve adjacent code, that's how PRs blow up

4. goal-driven execution: define success criteria upfront, loop until verified. without them Claude either loops forever or stops too early

5. use the model only for judgment calls: classification, drafting, summarization, extraction. NOT routing, retries, status-code handling, deterministic transforms. if code can answer, code answers

6. token budgets are not advisory: per-task 4000, per-session 30000. by message 40 of a long debug, Claude is re-suggesting fixes you rejected at message 5

7. surface conflicts, don't average them: two patterns in the codebase? pick one. Claude blending them is how errors get swallowed twice

8. read before you write: read exports, callers, shared utilities. Claude will happily add a duplicate function next to an identical one it never read

9. tests verify intent, not just behavior: a test that can't fail when business logic changes is wrong. all 12 of Claude's tests can pass while the function returns a constant

10. checkpoint every significant step: Claude finished steps 5 and 6 on top of a broken state from step 4. nobody noticed for an hour

11. match the codebase conventions: class components? don't fork to hooks silently. testing patterns assumed componentDidMount, hooks broke them without surfacing

12. fail loud: "completed successfully" with 14% of records silently skipped is the worst class of bug. surface uncertainty, don't hide it

what actually compounds instead of the next framework:

- the CLAUDE.md file as institutional memory across sessions
- eval-driven changes, not vibe-driven
- checkpoints over speed
- explicit conflicts over silent blending
- discipline over framework, every time
- one repo, one rules file, no exceptions

you don't need a better AI

you need better context engineering

complete playbook below ↓

[Pranjal Sahu @pranjalsahu]

People Who Grew Up Wealthy Are Sharing The Moment When They Realized They Were Privileged, And It's Eye-Opening

Funny !

https://t.co/QJCEr8DP2l

[Gabriele Berton @gabriberton]

It's cool to have your paper being called "the wildest CVPR 2026 result"

Also pretty good short summary of the paper

[Pranjal Sahu @pranjalsahu]

Masked-Diffusion Autoencoders for 3D Medical Vision Representation Learning
https://t.co/NsLg77M6eF

[Pranjal Sahu @pranjalsahu]

VoxTell: Free-Text Promptable Universal 3D Medical Image Segmentation

https://t.co/gG5jjSqefY

[Pranjal Sahu @pranjalsahu]

Data-free continual learning for medical SSL via model inversion.
https://t.co/471WFWZSvU

[Tech with Mak @techNmak]

Learn LLMs from Stanford this weekend.

Stanford's Autumn 2025 Transformers & LLMs course is fully public and 100% free.

In 9 lectures, you’ll unlock the exact mechanics behind:
• Flash Attention (3x faster models)
• LoRA (90% cheaper fine-tuning)
• Mixture of Experts (Massive efficiency scaling)

➕What's covered:

➡️Lecture 1: Transformer Fundamentals
→ Tokenization and word representation
→ Self-attention mechanism explained
→ Complete transformer architecture
→ Detailed implementation example

➡️Lecture 2: Advanced Transformer Techniques
→ Position embeddings (RoPE, ALiBi, T5 bias)
→ Layer normalization and sparse attention
→ BERT deep dive and finetuning
→ Extensions of BERT

➡️Lecture 3: LLMs & Inference Optimization
→ Mixture of Experts (MoE) explained
→ Decoding strategies (greedy, beam search, sampling) → Prompting and in-context learning
→ Chain-of-thought reasoning
→ Inference optimizations (KV cache, PagedAttention)

➡️Lecture 4: LLM Training & Fine-tuning
→ Pretraining and scaling laws (Chinchilla law)
→ Training optimizations (ZeRO, model parallelism)
→ Flash Attention for 3x speedup
→ Quantization and mixed precision
→ Parameter-efficient finetuning (LoRA, QLoRA)

➡️Lecture 5: LLM Tuning
→ Preference tuning
→ RLHF overview
→ Reward modeling
→ RL approaches (PPO and variants)
→ DPO

➡️Lecture 6: LLM Reasoning
→ Reasoning models
→ RL for reasoning
→ GRPO
→ Scaling

➡️Lecture 7: Agentic LLMs
→ Retrieval-augmented generation
→ Advanced RAG techniques
→ Function calling
→ Agents
→ ReAct framework

➡️Lecture 8: LLM Evaluation
→ LLM-as-a-judge overview
→Best practices and benefits
→Biases and pitfalls

➡️Lecture 9: Recap & Trending topics

From Stanford Online:
Rigorous instruction. Latest techniques. Free access.

Perfect for:
→ ML engineers building with LLMs
→ AI engineers understanding transformers
→ Researchers working on language models
→ Anyone learning beyond API calls

This weekend: learn the techniques that separate good engineers from great ones.

(I will put the playlist in the comments.)

♻️Repost to save someone $$$ and a lot of confusion.
✔️Follow @techNmak for more AI/ML insights.

[Pranjal Sahu @pranjalsahu]

Forehand Grip #shorts
Excellent tip to avoid wrist pain due to tight grip.
https://t.co/5Te0m3P1MV via @YouTube

[Pranjal Sahu @pranjalsahu]

@docmilanfar It is like saying, someone has already won the Nobel prize (ref: black hole work) and listing other (smaller) awards/achievements is like a waste of time.

[Veritas Daily @VeritasDaily]

@fermatslibrary

[Tech with Mak @techNmak]

Microsoft made 100B parameter models run on a single CPU.

bitnet.cpp: The official inference framework for 1-bit LLMs.

The math behind 1-bit LLMs is what makes them revolutionary.

Traditional LLMs use 16-bit floating point weights. Every parameter is a number like 0.0023847 or -1.4729.

When you run inference, you multiply these floats together. Billions of times. That's why you need GPUs, they're optimized for floating point matrix multiplication.

BitNet b1.58 uses ternary weights: {-1, 0, 1}.

That's not a simplification. That's a fundamental change in the math.

When your weights are only -1, 0, or 1:
→ Multiply by 1 = keep the value
→ Multiply by -1 = flip the sign
→ Multiply by 0 = skip entirely

Matrix multiplication becomes addition and subtraction.

No floating point operations. No GPU required.

This is why bitnet.cpp achieves:
→ 2.37x to 6.17x speedup on x86 CPUs
→ 1.37x to 5.07x speedup on ARM CPUs
→ 71.9% to 82.2% energy reduction on x86
→ 55.4% to 70.0% energy reduction on ARM

The speedups scale with model size. Larger models see bigger gains because there are more operations to simplify.

A 100B parameter model running at human reading speed (5-7 tokens/second) on a single CPU.

That's not optimization. That's a different paradigm.

Why 1.58 bits? Because log₂(3) ≈ 1.58. Three possible values = 1.58 bits of information per weight.

The key insight: These models aren't quantized after training. They're trained from scratch with ternary weights. The model learns to work within the constraint. No precision loss. No quality tradeoff.

[Alex Finn @AlexFinn]

The velocity in which Anthropic ships is unlike anything I've ever seen

I'm sorry, but we have to assume they have access to an AGI like model that nobody else in the world has correct?

They are simply holding it back so they can ship industry changing updates daily?

[Pranjal Sahu @pranjalsahu]

Random Forest in Machine Learning: Easy Explanation for Data Science Int... https://t.co/SvZEg2ypXm via @YouTube

[Pranjal Sahu @pranjalsahu]

VS Code Agent Mode Just Changed Everything https://t.co/DAl51L3fC1 via @YouTube

[Pranjal Sahu @pranjalsahu]

Python Tutorial: AsyncIO - Complete Guide to Asynchronous Programming wi... https://t.co/iBpfclKKMR via @YouTube

[Pranjal Sahu @pranjalsahu]

https://t.co/J2nuUpy2Tq

[Pranjal Sahu @pranjalsahu]

RAG Explained For Beginners https://t.co/eZ0jRY20Vn via @YouTube