Nitya

nityas notes dot com a place where I will keep track of what I’m reading and writing

@JehanGodrej Wow

had a good past week of training

doing things people said you can’t is important the tasks we use agents for should be targeted towards the increasingly ambitious, this requires using your brain

I really like this post and think it’s illuminating. I’ve felt a bit nervous about bun as a project since it seems buggy and not as stable/correct. The speedups don’t materialize as they take random micro-benchmarks out of context. But they tell a good story I guess, it’s ideas like this that will last For actual dev I’ve stuck to pnpm, node, vite/rolldown (also in rust), which I think are great options and I can’t complain

@charliermarsh I can’t tell if you are just successfully rage baiting here

I don't even know where to start with 5.6 Sol. I feel like I have superpowers. For the first time in my life I'm waking up naturally at 6am because I'm so excited to keep building with it. (This is 100% serious.)

@StasBekman +1 to other replies!

As I'm working on the Machine Learning Engineering and the Art of Debugging books, which I tremendously enjoy doing, I have been wondering if people actually care to read those. Most likely my content will be used as SKILL.md for AI -they will still be useful to the community.

Memory cost and capacity are significant issues for AI accelerators. Unlike game rendering, model inference can have a deterministic memory access pattern. You don’t need “random access memory” at all for model weights, and you could tolerate cold-start latencies in the multiple milliseconds, as long as continuous reads were delivered at the necessary bandwidth. NAND flash is over 100 times cheaper per GB than HBM, so there should be opportunity there, even after giving a flash controller a 1024 bit interface with HBM bandwidth. You could make a specialized pin protocol that just supported pipelined transfer of full 16KB+ pages from the flash to program-managed accelerator scratchpad memory and improve per-pin performance over HBM, but it might be more convenient to make it still look like a true random access memory with very fragile performance characteristics, where anything but sequential reads falls off a 1000x+ performance cliff. That has the advantage of automatically using existing cache hierarchies, and providing a natural path to update the flash memory with new model weights. With the stream-to-scratch interface, code has to be completely rewritten before it works at all, while the ram-emulation interface will start off just extremely slow, and you can incrementally sort out the changes for full performance. There may be cases where there isn’t enough scratchpad SRAM to hold the weights for a layer, which might force you to deploy the old optical drive optimization technique of duplicating data in multiple places on a sequential read to avoid seeking, but there would be capacity to burn. It might be possible to do something like cuda graph capture to record a memory access trace and have everything magically remapped to a linear sequence, but deploying programmer / agent elbow grease to manage transfers and access in a scratch ram ring buffer would be lower risk. A split memory system consisting of some channels of flash and some channels of HBM will probably be suboptimal compared to a uniform memory, but it could be much cheaper, and allow much larger models to be run. I think th case is strong for inference, but you have to stretch more for training. You can still linearize all the weight memory accesses, both reads and writes, but flash memory would quickly wear out from the writes, even if they were all perfectly page aligned. Replacing low-latency HBM with massively parallel cheap(er) DRAM at high latency might still be a worthwhile cost savings.

@paularambles it makes me sad but at every turn biotech is cursed as an industry tbh shockingly low pay, bureaucracy gets worse every year, global competition is winning, both being blocked from using any frontier models bc of safety but fake progress is hyped up

$65k to help build the future of biology in san francisco the first experiment is whether you can survive on that

@willccbb lfg congrats!!

we raised $130M @ $1B for our series A >$100M run rate we’re just getting started

@nicole_ripka wooo congrats!

Achieved some goals recently: Learned to drum ✅- 10 months of weekly sometimes 2x a week lessons, can kind of actually play now and get creative with beats Learn to dj & dj a real party ✅- months of practice with crossfader and YouTube and 10+ private lessons, just djed my friends wedding in Scotland Run a half marathon ✅- long runs on saturdays to train, never ran more than 5 miles at a time before this If you think you can or you think you can’t you’re right has always been my fave quote true for anything. Goals best met for me when I set deadlines, find reasonable tutors who I can build deep and meaningful relationships with and who push me.

@michlimlim @BoomPopHQ Congrats Michelle!

Introducing Flint Ads Agent: we optimize your Google Ads spend for you. Our agent figures out what’s costing you conversions, using our comprehensive data and optimization engine. Then, it executes the fix for you, and learns from the results. @BoomPopHQ 10x’ed conversions with it.

@base wooo @Must_be_Ash congrats!

this onboarding idea from @ExaAILabs is brilliant, get the user to complete your setup via offering credits i generally skip onboardings but this worked on me

thinking about the time I almost threw up after running 2 miles on the west side high way a few years ago

woke up on a Friday and ran a half marathon with my friends

@zebriez @turbopuffer I really enjoy their blog and am curious about writing culture at turbo

Worklife Update: I’m spending the summer with @turbopuffer 🐡☀️ This hardcore whimsical team (of only 38!) is doing something special and their product is exploding. The plan is to get even more people puffing and write about the experience. What do you want to learn about this mostly-bootstrapped pretty-Canadian AI darling making search cool again?

@Reshusaur @tinathetechie @liammatteson @ExaAILabs @browserbase ❤️

Thank you @tinathetechie @liammatteson and Song for hosting such a lovely dinner. Still feeling warm and fuzzy from all the laughter, conversations, and great food ❤️

@PradyuPrasad @rishabh16_ bell labs actively contributing to research in 1999 breaks the mental model I had of it, I thought they were shut down by then

@rishabh16_ @nityasnotes the OG en.wikipedia.org/wiki/LeNet

you can just name things after yourself

#when-we-cease-to-understand-the-world" target="_blank" rel="nofollow noopener">nityasnotes.com/bookshelf/#whe…

the first two chapters of this book are two of the best I read this year and then the author gradually started just entirely making things up

@sarahwooders @Railway @ExaAILabs love this sarah! realizing we met at a dinner a while back, full circle :)

Mods in Letta are pretty cool - agents can self-expand their capabilities. Today, I had our fully local finance agent (running on @Railway) install a mod to support web search with @ExaAILabs -- all via slack