Suman Karumuri

Anthropic's own study proves Vibe-Coding and AI coding assistants harm skill building. "AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average" Developers learning 1 new Python library scored 17% lower on tests when using AI. Delegating code generation to AI stops you from actually understanding the software. Using AI did not make the programmers statistically faster at completing tasks. Participants wasted time writing prompts instead of actually coding. Scores crashed below 40% when developers let AI write everything. Developers who only asked AI for simple concepts scored above 65%. Managers should not pressure engineers to use AI for endless productivity. Forcing top speed means workers lose the ability to debug systems later. ---- Paper Link – arxiv. org/abs/2601.20245 Paper Title: "How AI Impacts Skill Formation"

100%. If taxes in WA and CA are the same, moving to CA is a no brainer. The weather is nicer + better job opportunities.

Your website should improve itself. Sherpa gives you an entire conversion optimization team with one line of code. Here's how 🧵

Maturing is realizing that Tony Stark was a vibe-coder.

@lateinteraction What type of search are you referring to here?

judging by emails, seemingly every other lab is trying to hire leads for their search teams now; it kind of feels late for that?

@CitizenAgim @dougboneparth @grok identify this painting.

@dougboneparth When the second edible finally hits:

At some point in your life you will eat an edible and think it’s not working. It’s very important that you do not eat a second one.

Fascinating performance design direction: GenAI specializing your DB to your workload. This is fascinating to me due to some key parallels: * I already practice "eval-driven AI coding loops", and database performance is one of the happiest cases for this - clear conformance suites in correctness and clear goal to hill climb via optimization. * Moving this idea closer to the use case and runtime, while still staying safely offline, are both cool * In GFQL, our open-source GPU graph query language, we have started stacking specialization layers in the engine, where we take advantage of another phrasing of this: pay-as-you-go semantics. Simpler base layers can ignore complications of fancier language features so simpler queries can go faster, and fancier features get case-split so we optimize their different scenarios. Instead of one path straddling all cases, we have specialization layers and pockets. It used to be a LOT harder to identify the scenarios, refine the solutions, and build the safety & maintenance layers for this, while now we can easily adapt big conformance suites and run all sorts of test amplifiers whenever we add a new specialization. Overall, this paper leans into using AI to handle more complexity in, for now, narrow domains. It realizes databases are the happy case of easy for AI to test & verify, and the space of many optimizations that are well-understood: these make it a lot easier for eval-driven AI coding loops to hill climb up the performance charts for arbitrary workloads. Link:

me and claude code all day every day

Old world software estimates: 2 weeks (actually takes 4 weeks) New world software estimates: 2 weeks (actually takes 4 weeks, or 20 minutes)

@valyala @VasiliyZukanov Abstractions saves tokens. From what we have seen so far whatever abstractions that help humans also helps ai save tokens. Sometimes abstractions hide optimizations, but they are useful more frequently than not.

@VasiliyZukanov There should be another question: what is the purpose of generating the code in any programming language? Why AI doesn't generate an optimized executable for the target platform?

Honest question: why choose Python for the backend in the age of AI? When agents write almost all the code, and human preferences become less important, wouldn't it make sense to optimize for performance, scalability and safety with languages like Java, C#, Go, etc.?

The 20$ codex plan is worth more than the $200 Claude Code plan.

@rosslazer We are.

Is anyone building observability tooling that's native to agents? From the infra ground up. Not just an MCP server or observability for agents themselves.

@EthanLipnik 👋 Early versions of Claude Code used RAG + a local vector db, but we found pretty quickly that agentic search generally works better. It is also simpler and doesn’t have the same issues around security, privacy, staleness, and reliability.

@JustJake Git at scale was never a solved problem. FAANGs have moved to git replacements even before agents were a thing. The biggest issue scaling git was doing it in a way that maintains backwards compatibility. CI/CD infra seems to be real bottleneck to me.

If you haven’t done this already It’s going to get very, VERY painful very VERY soon

VC vs. founder

@ankurnagpal Given the number of online scams, I found the best way for the govt to contact me is via physical mail.

REAL QUESTION How bad is it if I ignore all physical mail from today onwards Like throw it straight in the trash Assume anything important enough will find another channel, yes?

Average iops per gigabyte of storage (queue depth 1): 40 years ago (HDD): ~1500 Today (HDD): ~0.01 Today (SSD): ~10 Today (eMMC): ~3 Storage is much cheaper, but also much further away.

A must-bookmark for vibe-coders. @YCombinator’s guide to making the most of vibe coding:

@kellabyte Disaggregated storage(separate compute from storage) doesn’t have to eschew local storage. In Kaldb, we use S3 for durability but download the data locally for fast queries. With thoughtful sharding we can spin up data compute nodes on demand to serve the data.

Not all customers need the advantages with the disadvantages of remote storage. Some people need the instant spin up and scalability of compute and all the cool features remote storage enables. Some ppl benefit their cutomers way more w/ p99’s being lower 24/7.

I don’t think this take will age well. The industry keeps separating storage and compute then opening the door to competitors to bring them back close together. I think best DB companies will allow you to config local low latency storage or config benefits of network storage.

This killed me.