AI researcher x Founder | prev @observeAI (YC’18), @hexoai & 4x other YC startups | Built & launched 6 times last year • i will win regardless
Joined December 2021
- Tweets5.1K
- Following977
- Followers12.2K
- Likes6.1K
Yashraj Shukla retweeted
Huge congrats to @warpsciences for coming out of stealth today. We're proud to be early supporters (more on that later).
The mission is to train AI to do chemistry and to use that to reestablish American medicinal chemistry supremacy.
@kevinweil + @dpatil + Sol Reisberg + @manohar_paluri + their early hires are a dream team to take this on. Legends in the Valley all with personal origins in science.
💥 Introducing Warp Sciences, a new company applying AI and robotics to help the world develop new medicines faster.
We’ve raised a $215M Series A led by @lightspeedvp to bring frontier AI out of the browser and into the lab, and build scientific capacity here in America.
I co-founded Warp with my long-time friend @dpatil, FAIR co-founder @manohar_paluri, and drug hunter Sol Reisberg. The early team is full of incredible people like @bkspears9, who helped achieve Q>1 fusion at LLNL.
Yashraj Shukla retweeted
Me and my friend @whynesspower have been doing this for more than half a decade now.
We have tracked every single hour of work we have done for the last 5+ years.
We were hyper obsessed over maximizing the work we could do, we essentially treated ourselves as athletes.
Finally it's going mainstream.
introducing Promethee, an app that turns your work into a multiplayer game, to perform.
now available for everyone.
promethee.io
Yashraj Shukla retweeted
This would be a perfect RL environment.
Take real shipping data, turn real logistics problems into verifiable tasks, and let models learn to optimize routes, schedules, inventory, and the entire network.
Basically, logistics software as a strategy game for RL.
It’d be awesome if logistics software could be visualized and turned into a strategy game.
You manage transportation, routes, and the entire logistics network like you’re playing a strategy game.
Inspired by @DilumSanjaya, I also used Opus 5.5 to build it.
Yashraj Shukla retweeted
Archgen Labs (@archgen_) is now backed by Y Combinator (@ycombinator)
We are a research lab for silicon design, with one ambition: make chip design 1,000× faster.
For decades, computing advanced by shrinking transistors. The next chapter for compute workloads will increasingly depend on building vertically - stacking active logic layers to create 3D chips.
This makes designing those chips much harder but also opens up new possibilities for performance and efficiency.
At @archgen_ , we’re bringing together foundational models, EDA tools, and physics-informed surrogate models to explore more designs, predict their behavior faster, and optimize how they’re built.
Grateful to Kulveer Taggar (@kul), Garry Tan (@garrytan) and the @ycombinator team for backing us, and to the customers, mentors, and collaborators who have helped us get here.
If you want to get to tapeout faster, we’d love to talk. DM us.
#Semiconductors #OpenROAD #OpenLane #OpenSourceEDA #ChipDesign #VLSI #PhysicalDesign #AIForHardware #3DIC
@naveen_venk
Yashraj Shukla retweeted
I ported Google's TimesFM 3.0 to run natively on Apple silicon.
TimesFM is a 330M parameter foundation model for time series forecasting. It reads history in patches of 32 points and writes the next 64 in a single forward pass, nine quantiles per step. Until now you needed PyTorch. mlx-tsfm runs it on MLX, Apple's array framework, with no torch install.
Numbers on an M4 Max, 512 point context, 64 step horizon:
- 12.5 ms per series at batch 1
- 641 series per second at batch 32
Max difference from the PyTorch reference: 2.1e-6, which is float32 noise
The 1.9x speedup at batch 1 came from mx.compile, not quantization. At 330M params this model is dispatch bound, so bf16 and int8 were both slower than fp32 with a compiled graph. That surprised me. It is why the default stays fp32.
Two images below. The first is for anyone who has never touched a forecasting model. The second is the decode path with tensor shapes, if you want to check my work.
Code is MIT. Weights are under Google's research license.
github.com/rachittshah/mlx-t…
Introducing TimesFM-3, a state-of-the-art time series foundation model that enables accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models across major benchmarks.
More on the blog →goo.gle/4x5WGpD
ALT TimesFM-3 architecture.
Yashraj Shukla retweeted
One commonality between Nikesh Arora and Dara (uber) is that both are pretty open about not being 0-to-1 guys
Their superpower is taking ginormous companies and somehow keeping them moving fast, with less bureaucracy than you’d expect
hard to imagine the founders of Palo Alto Networks doing 40+ M&As in 5 years
Nikesh Arora is a prime example of why you should never let a title intimidate you.
When he was named CEO of Palo Alto Networks, he was called a "controversial hire."
He had never worked in cybersecurity. Never been a public company CEO. Never sold enterprise software.
And even today, he says he still deals with imposter syndrome.
"I'd never done cybersecurity in my life. This is a cybersecurity company. I'd never been a public company CEO. This is a public company CEO job. I'd never sold enterprise. I was a consumer guy."
"Over time, I began to learn."
Yashraj Shukla retweeted
a quite delightful reminder that every boring thing in your day was somebody's life's work. music you can play whenever you want was once someone's idea of the highest human happiness. same with hot water
"it is to humanity’s credit that we remain restless in the midst of all of this progress. we continue to look forward, pushing the frontier further...but our lives today are a gallery of past generations’ heroic efforts to do the same."
ordinaryabundance.com/
This might be the most beautiful thing I've read in my life. What a privilege to be alive in 2026.
ordinaryabundance.com
Yashraj Shukla retweeted
Prime Agent is a general-purpose coding harness
On ARC-AGI-3, it scores 95.5%, surpassing the human-expert baseline, but the gain is not benchmark-specific.
We see major improvements across models when compared to their proprietary harnesses:
Yashraj Shukla retweeted
What a wild world we’re living through. I wrote about the Series A squeeze recently. Here’s what else I’m seeing from my seat on the investing side..
1) VC has effectively fully become vibes capital. The market reality at the early-mid stage is completely divorced from fundamentals. I can go longer with even a resemblance of a straight face predict what happens to a company when they go to fund raise.
Some rounds are insanely wild with nothing to show, some rounds that seem like a sure shot are struggling to raise the capital they need. If you’re the “in vogue sector”, you’ll be just fine. This has been happening for some while but it’s fully banana town now. And unlike most who think a correction happens I think this continues for 12-18 months (dry powder, AI tailwind) minimum. The most common phrase in VC land right now is “you gotta play the game on the field”. This is a spiral that continues to keep the volatility and prices high.
2) Whereas on the public markets it’s also equally bananas. It used to be in the meme stock era that you’d see daily swings of >5% (+\-). But now even trillion dollar stocks are not immune to swinging that much based on vibes and model releases. Memory and the KOSPI index are recent examples and this will happen to other sectors soon. Rotations are getting faster and faster.
In the long run, the same fundamentals of building a great company with control of your own destiny (profitable, clean cap table etc) will end up just working out fine. But in the short term, in this vortex of hyper competition where capital is needed as a weapon, you should make sure you understand what’s going on before you dip your toes in the capital markets.
No matter what it’ll be interesting to see how this shakes out when we look back in this era. Obligatory happy to chat if helpful!
Yashraj Shukla retweeted
I think there's a lot of confusion and misinformation about Zhilin's visa, H1B lottery, immigration, and why he left the U.S.
Zhilin had many opportunities to stay in the U.S. if he wanted to. In fact, I was on an email thread with a senior Apple exec (Tim Cook's reports) asking whether Zhilin would consider joining Apple. I said that he wanted to go back to his homeland. The response was, well, we have an office in Beijing if he'd like to join there.
But Zhilin was quite determined to go back and build a startup. I remember him telling me that if he didn't at least try starting his own company, he would regret it for the rest of his life. I respect that, and he was right.
Many of my international PhD students do choose to stay in the U.S. Of course the U.S. immigration process can be quite intimidating and uncertain, even for superstar PhD graduates from places like CMU.
Yashraj Shukla retweeted
Stats which made YC give us an interview this time:
(on our 3 month old product)
>$200k cARR
>$20k revenue
closed two F500 global, four more in pipeline
The 10 min call went like bullet and they loved the idea. Too bad could not show enough US traction
Untill next time? 💪
Yashraj Shukla retweeted
AI is going to shift humanity’s focus from Survival Mode to Creative Mode.
Cognition CEO @ScottWu46:
“ We've been spending all this time living in Survival Mode as a species. Now we're going to be living in Creative Mode.”
”Minecraft Survival Mode is where you're growing food, you're making sure you're safe from the monsters at night.”
“Creative Mode is—everything's up to you. You have all the resources at your disposal. If you want something to happen, it'll happen. The only question for you is: what do you want to make happen?”
My conversation with @ScottWu46, founder and CEO of @Cognition, the company behind Devin, the first AI software engineer.
0:00 Scott Wu's Obsession With Winning
2:06 Competitive Programming, Games And Finding His People
4:24 Family, Go, And The Roots Of Scott's Competitiveness
8:35 Why Losing Hurts More Than Winning Feels Good
9:38 What Winning With Devin Looks Like
12:55 Devin Today: The AI Software Engineer
13:52 Software As The Human-Computer Interface
18:45 Why AI Progress Is Hard To Intuit
20:39 Thinking About AI From First Principles
22:57 What Happens When Agents Can Work For Months
30:18 The Original Thesis Behind Cognition
31:12 Launching Devin And Handling Criticism
37:17 Finding Product-Market Fit In The Enterprise
42:41 How Cognition Deploys Devin Inside Large Companies
48:34 Measuring ROI Instead Of Token Spend
50:01 Why Cognition Wants To Be Model-Neutral
52:18 Why Focus Lets Startups Beat Giants
57:14 Independence, Acquisitions, And Building A Generational Company
1:00:27 Why Money Is Not The Goal
1:03:42 One Life: Going For It All
Includes paid partnerships.
Yashraj Shukla retweeted
The opposite of paranoia is pronoia, the irrational belief that the universe is conspiring in your favor. I have built my entire career on pronoia and I highly recommend it.
Yashraj Shukla retweeted
Mark Cuban just described the largest wealth transfer of the AI era.
Almost nobody understood what he said.
Cuban: “There are 33 million companies in this country. Aren’t going to have AI budgets. Aren’t going to have AI experts.”
Not tech startups.
The shoe store. The regional trucking outfit. The accounting firm with 12 employees.
The businesses that actually run the physical economy.
They know AI is coming. They have no idea what to do with it.
Cuban: “You’ve got the head of Microsoft saying software is dead because everything’s going to be customized to your unique utilization.”
Software is dead.
The SaaS era ran on one rule. Build a generic product. Force millions of companies to bend their workflows around it. Charge rent forever.
AI ends the contract.
The business stops bending to the software. The intelligence bends to the business.
But customized by whom.
The third-generation manufacturer cannot tell Claude from Gemini. The county hospital is staring at a reactor asking where the light switch is.
Cuban: “Who’s going to do it for them?”
That question is worth more than the frontier models themselves.
Hundreds of billions are being burned to build the foundation. The smartest engineers alive are locked in a bloodbath over who owns the base layer.
Let them fight.
Let them burn the capital. Let them drive the cost of raw intelligence toward zero.
Because the wealth does not collect where the brain is built.
It collects where the brain meets the business.
Every ambitious kid in college right now thinks survival means a seat at OpenAI or Anthropic.
Cuban is staring at the other 99 percent of the economy.
Learn the models. Then learn the messy, unglamorous reality of how a 50-person company actually operates.
Walk through the door. Understand their problems. Wire the intelligence directly into their revenue.
That is not a job title. That is an entire economic class being born.
You do not need to build the brain. You need to build the nervous system.
The biggest winners of the electricity era were not the engineers who built the generators. They were the ones who walked into dark factories and showed the owners where to plug in.
33 million companies are standing in the dark right now.
Silicon Valley is racing to build the god. The fortunes will belong to whoever teaches him a trade.
Yashraj Shukla retweeted
using context graphs to find best gift for parent's 28th marriage anniversery. ~10 days left, but too lazy to sit and think. so index'ed transcripts of >50 hours of call recordings with my mom and dad into the context graphs sitting on top of supermemory and graphiti.
now i let claude 4.6 query it (enterprise, zero retention plan). for mom, it suggested me to buy the most no brainer thing of all: gold bangles🤦🏻♂️
what it recommended me for dad blew my mind - a trekking tent, could not even think of it in a million years, but makes so much sense, he probably mentioned it years ago that he struggled in an overnight camping trip. this technology is surreal
Yashraj Shukla retweeted
As the famous quote by Phil Karlton goes: "There are only two hard things in Computer Science: cache invalidation and naming things."
Well, at least solved one. The Cache Invalidation
Before this, if you wanted to build a query cache for a vector database that handles both reads and writes, you faced a mathematically brutal dilemma. If an application executes a search query and caches the top-k results, and then milliseconds later a new vector is inserted into the database, does the cache remain valid?
Because vector search is based on proximity in continuous space, a single new vector could theoretically alter the top-k results of almost any previously cached query.
@kyrodb solves this by finding a mathematically rigorous way to perform Targeted Invalidation without needing to fully compare the newly inserted vector against every single cached query.
It does this through Geometric Prefix-Bounding:
When a new vector x is inserted, KyroDB doesn't flush the cache.
Instead, it looks at the first 32 dimensions of the new vector and the cached query q.
Using the Cauchy-Schwarz inequality, it calculates a mathematically proven upper-bound (for cosine) or lower-bound (for Euclidean) distance.
If this bound proves that vector x cannot possibly beat the worst distance in the cached top-k result, KyroDB safely skips invalidation.
Is this a big deal? Well, yeah
It gives you the best of both worlds, which was previously thought to be an impossible tradeoff in vector DBs:
→ Strict Correctness: You get exact, read-your-writes consistency. No stale data.
→ Massive Performance: You preserve the cache hit rate for the vast majority of queries that aren't geographically near the new insert, without doing full, expensive distance recalculations.
Yashraj Shukla retweeted
Karpathy's autoresearch loop (write code, train for 10 minutes, check if it improved, keep or discard, repeat) is getting cloned in other verticals to improve models. Bitmind tshipped a deepfake detection toolkit built on the same chassis: point an agent at 60+ image datasets overnight, wake up to 50 iterated experiments and a competition-ready model. Thanks to Karpathy, this sort of autonomous ML research is becoming an essential primitive for frontier AI teams.
Agents are the future. Inspired by autoresearch and arbos we released our deepfake research training toolkit: DFResearch: github.com/BitMind-AI/dfrese…
Experiment autonomously to train the best deepfake detection models. Integrated to download data, output submission ready-results, and full guide to adding custom models, datasets.
Yashraj Shukla retweeted
DeepSeek just dropped a new paper — and it's not about a new model.
It's about the infrastructure bottleneck that limits every agentic LLM at scale. And their fix nearly doubles throughput.
DualPath (arxiv.org/abs/2602.21548): the bottleneck in multi-turn agentic inference isn't compute anymore. It's storage bandwidth.
When agents run at scale, every new turn reloads the full KV-Cache from external storage. At long contexts + high concurrency, this I/O completely dominates. The GPU sits waiting on the NIC.
In standard disaggregated setups (prefill + decode engines), there's a brutal asymmetry:
— Prefill engine NIC: saturated loading KV-Cache
— Decode engine NIC: sitting completely idle
The bandwidth is right there. It's just not being used.
DualPath adds a second loading path:
Path 1 (existing): Storage → Prefill engine
Path 2 (new): Storage → Decode engine → Prefill engine via RDMA
The idle decode NIC now does real work. RDMA transfers over the compute network avoid congestion without touching latency-critical inference traffic. A global scheduler dynamically balances load across both paths in real time.
Results on 3 models with production agentic workloads:
📈 Offline throughput: up to 1.87×
📈 Online serving: 1.96× average — without violating SLO
As agents get more capable — longer context, more tool calls, persistent memory — KV-Cache I/O only gets worse.
DeepSeek is solving this at the infrastructure layer while everyone else is focused on model benchmarks.
One of those "obvious in hindsight" ideas: the bandwidth was always there.
Yashraj Shukla retweeted
n8n is dead. I fed my AI all my workflow JSONs and it replaced every single one.
I ran n8n for over a year. 15+ workflows on a Hetzner VPS. HTTP nodes, webhooks, JSON parsing, conditional branches, error handling.
Then I did something stupid: I exported every workflow as JSON, dropped them into @OpenClaw, and said "replicate all of this."
It took 20 minutes.
Not 20 minutes per workflow. 20 minutes TOTAL. The AI read every JSON, understood the logic, and rebuilt each one as a self-healing cron job.
Here's the difference nobody talks about:
n8n workflow:
→ 47 nodes connected with spaghetti
→ API changes? Rebuild 6 nodes manually
→ Edge case? Add more conditional branches
→ Breaks at 3am? Wake up and debug
OpenClaw workflow:
→ "Scan X for trending AI news, write a 1000-word article, generate a thumbnail, publish to WordPress, submit to Google Search Console"
→ API changes? The AI adapts automatically
→ Edge case? It figures it out
→ Breaks at 3am? It reads the error and tries a different approach. You sleep.
The craziest part: I now have 20 autonomous jobs that do MORE than my n8n ever did.
Because here's what n8n can't do:
• Write a full SEO article from scratch
• Generate images for thumbnails
• Decide NOT to publish if the content is trash
• Review its own mistakes every 4 hours
• Learn from yesterday's failures
• Generate full YouTube videos — script, voiceover, stock footage, animated whiteboard, editing — completely autonomously
That last one blew my mind. I give it a YouTube URL, and it comes back with a fully edited video summary: AI-generated narration, Excalidraw animated scenes, relevant stock footage, transitions. No editing software. No templates. The AI figured out how to use Remotion, FFmpeg, and TTS on its own.
Try doing THAT with n8n nodes.
My AI scrapes Instagram, transcribes reels, hunts expired domains, deploys websites, writes posts on X — all on cron. If one task fails, it self-corrects and keeps going.
n8n doesn't think. It executes.
OpenClaw doesn't just execute. It understands.
You can literally hand it your n8n JSON export and say "do this but better." And it will. Because it doesn't need nodes and connectors. It needs intent.
My n8n VPS is still running. I haven't logged in for 3 weeks.
Stack: Mac Mini, @OpenClaw, plain English prompts. That's it.