GP at @Venrock - Helping Founders, ex CPO @confluent, ex VP Eng @uber, ex @LinkedIn
Joined July 2009
- Tweets77
- Following205
- Followers282
- Likes83
Ganesh Srinivasan retweeted
Today is a hard day. I shared this note with the @linear team today: We’ve made the difficult decision to increase our workforce. This is not a cost-cutting exercise or a reflection of anyone’s performance. We’re simply reimagining every role for the agentic AI era. We’re hiring. We’re sorry about that.
Great launch from the Oumi team!
🚨 The era of general-purpose AI is over.
Today we're launching Oumi. 🚀
The platform that lets any team build custom AI models — in hours, not months.
Just describe what you need. Oumi builds it. #VibeML
Higher quality. Lower cost. Fully yours.
Ganesh Srinivasan retweeted
New benchmark results: Newton beats larger frontier models on Physical AI object reasoning while remaining small enough for real-world, real-time deployment.
This proves our core thesis at @PhysicalAI: making sense of the physical world requires spatiotemporal reasoning - natural for humans, still a frontier challenge for AI.
Ganesh Srinivasan retweeted
If AI isn’t truly open, it will fail us. We can’t close in a black box our greatest invention yet just so that a few can freely monetize. AI needs its Linux moment, and so we started working towards it. This can only succeed if we all work together!
#oumi #opensource #collaboration
@rsalakhu @svlevine @larry_heck
@karpathy @atalwalkar @prfsanjeevarora @tsvetshop
@hhexiy @sainingxie @larry_heck @AnimaAnandkumar
@JunjieHu12 @georgiagkioxari @profjoeyg @pliang279
@danqi_chen @ChrisGPotts @BillMacCartney @vinodv
@tur_gokhan @dilekhakkanitur @j_foerst @gingsmith23 @kahinish @jamesjoaquin @gan3sh @ethanjb @kirbywinfield @egonzdp
Ganesh Srinivasan retweeted
While we celebrate @deepseek_ai 's release of open-weight models that we can all play with at home, just a friendly reminder that they are not *open-source*; there’s no training / data processing code, and hardly any information about the data.
Ganesh Srinivasan retweeted
Three of our co-founders, Jaime Lien, Nick Gillian, and @ipoupyrev co-authored the paper “Soli: Ubiquitous Gesture Sensing with Millimeter Wave Radar.”
This research shows how AI and non-visual sensors can create new ways to perceive and interact with the physical world. [1/3]
Ganesh Srinivasan retweeted
AI embedded LiDAR camera.
Context analysis with @PhysicalAI
iPhone's LiDAR
Adafruit's Feather as a BLE keyboard input
Built in Swift
Ganesh Srinivasan retweeted
Profile-guided optimizations (PGO) reduce capacity requirements and latency by 10-30% for Go, Rust, and other statically compiled languages in production. 🚀 All this without changing code.
If this sounds too good to be true, its not. Processors waste significant time on front end stalls; i.e., I-cache and I-TLB misses, and branch mispredictions. These stalls are at the micro-architecture level, so you won't see them in your tools unless you look at hardware-level performance counters.
Boost the performance of your Go services by 10-30% without touching a line of code!
Profile-Guided Optimization (PGO) uses runtime profiles to enhance your application's latency & efficiency.
Read more about it on our blog: gitar.co/blog/unlocking-spee…
Ganesh Srinivasan retweeted
I recently investigated the 1brc challenge and its Go implementation, especially in the context of PGO (Profile-Guided Optimization) which we upstreamed to Google. However, I found that the benchmark is predominantly I/O-bound, limiting the potential benefits of advanced compiler techniques like PGO. However, there is still 3.59+% improvement via PGO.
My analysis was conducted on a Mac OS with M2/64GB hardware. See details below:
1) Without PGO, average time per run = 8.08s
2) Profiling reveals that approximately 66% of the execution time is dedicated to I/O operations, specifically buffio.(*Scanner).Scan, indicating that it's an I/O-bound benchmark. Additionally, strconv.ParseFloat accounts for about 9% of the time, while accumulator::ensure contributes to 0.9% of the overall execution duration.
3) The Profile-Guided Optimization (PGO) we introduced in the Go compiler yields a 3.59% enhancement in this benchmark. This improvement is primarily due to the inlining of several critical functions across packages. Notably, when the inlining threshold was raised from 80 to 2000 for hot functions, both accumulator::ensure (100) and strconv.ParseFloat (1505) were effectively inlined, contributing to this performance boost.
// benchstat
pkg: github.com/warpstreamlabs/on…
│ before.txt │ pgo_after.txt │
│ sec/op │ sec/op vs base │
YourFunction-12 8.085 ± 2% 7.795 ± 1% -3.59% (p=0.000 n=10)
4) Another interesting finding is that by applying Profile-Guided Optimization (PGO) on slice sizes, currently manually set at 1<<5, we can entirely eliminate the accumulator::ensure function from the hotpath.
Increasing the slice size to 1<<9 ensures that there's no need to call ensure for expanding slice sizes [and the copies], though this is contingent on using the same measurements.txt file from run-to-run. Similarly, PGO can be applied to optimally set the buffer size in bufio.NewReaderSize(sr, 1<<19).
References:
1. 1brc challenge: teapaw.org/gunnarmorling/st…
2. Go implementation: github.com/warpstreamlabs/on…
3. Our PGO upstreamed proposal: go.googlesource.com/proposal…
📢 "The One Billion Row Challenge"
How fast can YOU aggregate 1B rows using modern #Java? Grab your threads, flex your SIMD, and kick off 2024 true coder style by joining this friendly little competition. Submissions accepted until Jan 31.
👉 morling.dev/blog/one-billion…
Why do mobile apps from extremely successful and profitable companies still suck? Here is a an endless loop of an eye test done by the myAT&T, I think the goal is to help me train my eye to spot random spinners on the screen!
Ganesh Srinivasan retweeted
We created Trunk, a marketplace for Postgres extensions. It powers Tembo Cloud and is growing fast.
Love Postgres extensions? Give Trunk a try: pgt.dev
Thanks for sharing Gunnar!
Ganesh Srinivasan retweeted
Not all 1k tokens are created equally.
The cost of GPT-4 32K is $0.12/1k tokens.
This is mostly a cost pass-through. This will have to change.
This will have to change from cost to value/1k tokens
The measure of value or revenue produced for every 1k tokens of output.
Because not all 1k tokens are valued equally.
Asking a model for a thanksgiving recipe, has a certain value associated to it.
Asking a model for help understanding why your engine has a loud hissing sound coming from the belts, has a certain value associated to it.
Asking a model to summarize 12 hours of security footage across 12 security zones, has a certain value associated to it.
Asking a model to help triage optimal growing conditions across an industrial greenhouse, has a certain value to it.
Not all 1k tokens are valued equally.
While this first era has focused on unlocking capabilities.
The next era will focus on driving value, measured per 1k token.
Decoupling infrastructure and compute from pricing.
Solving deeply important problems and challenges in our world. Where combining the knowledge of the digital world, with the entropy of the physical world, introduces real magic.
Ganesh Srinivasan retweeted
We are thrilled to announce our collaboration with @Infineon to accelerate the deployment of #PhysicalAI. Infineon's advanced sensors, coupled with our foundational model, aim to endow machines, devices, and environments with an unprecedented level of perception and intelligence!
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
Ganesh Srinivasan retweeted
🚀I am excited to reveal Nile to the world finally!
Nile is serverless Postgres database for modern SaaS
thenile.dev/
Nile virtualizes tenants into the database to provide
1. Tenant isolation and per-tenant backups
2. Drop-in organization and user management
3. Global tenant placement
4. Tenant-aware vector embeddings
5. Instant internal customer dashboards
6 Effortless elasticity
I had a great conversation with Ekanth on building systems that can handle scale. He has a very unique insights with his experience building out large scale data systems at companies like Uber and Notion that experiences massive growth.
🎙 On #TheOuterJoin, @gan3sh chats with @NotionHQ's Ekanth Sethuramalingam about why he joined Notion, his thoughts on building new primitives, and the challenges of building large-scale systems while experiencing massive growth. Listen below. vnrk.co/42OTbVf