Divij Vaidya

I tweet about life experiences (mostly tech related) | Opinions my own

[Matei Zaharia @matei_zaharia]

You can now query your lakehouse data in real time at <10ms latency! Any table, just faster, at 12,000+ QPS, on the existing open data formats. Powered by what @rxin's team affectionally called Reynold's Dream Engine (Reyden).

[Matei Zaharia @matei_zaharia]

Really excited to open source a new project: Omnigent, a meta-harness for AI agents.

It lets you build multi-agent coding and custom agents, sitting above Claude Code, Codex, Pi, and agent SDKs to let you compose them. It also adds live collaboration and rich control policies.

[Perplexity @perplexity_ai]

We're open-sourcing the Unigram tokenizer we rebuilt to reduce CPU utilization by 5-6x.

Small rerankers and embedders run in single-digit milliseconds on GPU, making CPU tokenization a meaningful share of total latency.

https://t.co/QUnHeiho56

[Matt Garman @mattsgarman]

Most people thought flat data center networks would never work at hyperscale.

The @awscloud team figured it out.

Resilient Network Graphs are a completely new network architecture, now live.

33% better throughput, 40% less network power.

https://t.co/LEaIItVqe9

[GDP @bookwormengr]

https://t.co/KHnPhxJBiz

[Shashank Kumar @shashank_kr]

We recently built an AI assistant inside @Razorpay called Slash.
It reads our entire codebase, debugs production incidents, reviews specs, writes code, reviews every single PR, answer tech queries and also raises PRs for small features.

It's easily accessible through Slack. We can tag it in any Slack thread, describe the problem in English, and it gets to work.

Six weeks ago, Slash handled 122 tasks in its first week. Last week it handled 14000+. Queries, analysis, bug fixes, PR reviews, test runs and work that earlier lived across scattered tools and teams can now be done with Slash right within Slack. 1000+ people used it in a single week because it got their work done faster. The whole adoption has been completely organic.

The numbers from last week have been very encouraging - 14,854 tasks completed. 2,150 PRs raised, 1,152 merged, 45% of those PRs shipped with zero human rework.

A payout gets stuck mid-retry during a live incident, an engineer tags Slash and within seconds, it cross-references logs with code and pinpoints a state machine bug blocking the retry-to-failed state transition. Tells the team exactly which logs to check and how to resolve the incident.

With its K8s analyzer skill, Slash scanned a single namespace, right-sized all 11 workers using 48-hour P95 pod metrics, and raised the PR. One run saved $560/month.

A marketing banner bug was fixed with few prompt iterations with a PR raised, merged to prod and deployed in minutes. No front-end developer touched the code.

Security teams ran static security testing and remediation through Slash at org scale. Thousands of findings were purged and many more got validated autonomously.

But Slash isn't just an engineering tool.

Account managers now trace stuck customer payments and integration failures through Slash instead of pinging engineers on Slack. L2 product support tickets get triaged by Slash before they reach engineering.

250+ non-engineers ran thousands of sessions last week. PMs used it for research on our payments infra, customer interviews and product features sometimes raising PRs of their own. Analytics teams built SQL pipelines. 11% of all sessions came from people outside tech and product.

On our company bakkar (watercooler) Slack thread, someone asked Slash jokingly to assign tasks to everyone and it responded in the same tone. It seamlessly started participating in inside jokes and conversations.

The quality compounds with use. Engineers who shipped 11+ Slash PRs averaged a 63% merge rate without rework. First-timers averaged 37%. Across the org, human review comments per PR have dropped more than 40% with Slash starting to do in-depth review of every single change.

We're still early. Large cross-repo refactors, fully agentic sdlc and plan mode are next. But Slash has already changed how people at Razorpay build, debug, and ship every day.

[Rohit Ghumare @ghumare64]

https://t.co/fRx8o2GYVs

[Aditya Kondawar @aditya_kondawar]

Rajeev Jain, CEO, Bajaj Finance, on Q3 Concall on the Impact of AI –

“AI listened to 2 Cr calls, converted voice to text, and gave us data. Text-to-data conversion happened for 5.2 lakh customers. As a result, we generated 100,000 new offers for which we did not have information earlier.

“This capability did not exist in Q1 and Q2. It just got deployed. We’ll be able to listen to 100 million calls next year,” said Jain. He added that loan disbursements through AI-powered call centres stood at about Rs 1,600 crore. That’s ~ 10% of the Rs 16,545 Cr of disbursals in Q3FY26

Data converting -- data from those calls led to another INR 325 crores of volumes. So, this is just our first attempt.

Over the next six months, Bajaj Finance plans to invest heavily in its agents. The company expects to have more than 800 autonomous agents across sales, operations, HR, IT, risk, and DMS in the next fiscal.
Similarly, in terms of 100% of videos are now generated by us using AI, 100% of banners are generated using AI, 2.7 lakh videos were generated, and 1.2 lakh banners were generated. At the customer engagement level, we have 11 AI text BOTs that are live that engage with the customer. So rather than sending dumb SMSs for 11 products now, we have an AI text BOT, which allows you to engage, interact, and respond to your queries.

The company has 26 products. All 26 will be live between April and May'26. So, there will be no communication that we'll be sending, which will not have a -- whether service or sales, which will not have a conversational BOT embedded in it.

At the branch and point of sale, existing customers face match that we're doing, we did 46 million face matches to ensure this is the same customer, if it's an ETB customer who had actually principally come in, giving us much better control over identity.

Customer onboarding in terms of document -- ensuring that auto-fill of the document happens, whether it's a PAN card or an Aadhaar. There are 43 such documents that the company has now mapped, which an image extracts with a 95% - 96% accuracy and populate data in our platforms, delivering significant productivity for our employees.

Auto quality check of documents is now 41%. As we sharpen the model, it will take us to between 85% and 90% over a period of the next 15-odd months

On technology development, we are clearly seeing between 25% - 45% efficiencies emerging in terms of the development process, depending on whether it's a legacy platform, then the benefit is much lower, or rather, I would say, none. But if it's a digital infrastructure, then the efficiencies can be as high as 45% - 47%. So significant work is being done.

Src – Q3 Concall, no reco

[Divij Vaidya @divijvaidya]

Key insight for me was the test harness tuning/optimization required to course correct without letting the agent run hours on testing

[Anthropic @AnthropicAI]

New Engineering blog: We tasked Opus 4.6 using agent teams to build a C compiler. Then we (mostly) walked away. Two weeks later, it worked on the Linux kernel.

Here's what it taught us about the future of autonomous software development.

Read more: https://t.co/htX0wl4wIf

[Anthropic @AnthropicAI]

New Engineering blog: We tasked Opus 4.6 using agent teams to build a C compiler. Then we (mostly) walked away. Two weeks later, it worked on the Linux kernel.

Here's what it taught us about the future of autonomous software development.

Read more: https://t.co/htX0wl4wIf

[David Marcus @davidmarcus]

A few thoughts about PayPal, nearly 12 years after I left.

I woke up this morning to dozens of messages from former PayPal colleagues. It pushed me to finally speak up.

I never spoke publicly about the company after I left. Part of that was loyalty to John Donahoe, who gave me an unlikely opportunity, handing the reins of PayPal to a startup guy who, on paper, had no business running a then 15,000-person organization. But part of it was something else: I had left. I chose not to stay and fight for the changes I believed in. Speaking from the sidelines felt like armchair commentary. Easy opinions without the burden of execution. So I stayed quiet.

But twelve years of silence is long enough. And today's news makes it clear the pattern I've watched unfold isn't self-correcting.

I left PayPal in 2014 because I was deeply frustrated. We had executed a silent turnaround of a company that had lost its soul. We brought back engineering talent, shipped good products quickly, and acquired Braintree and Venmo. The company was on a tear. So much so that Carl Icahn felt compelled to accumulate a position in eBay and push for a PayPal spinoff. At the time, eBay decided to fight Icahn.

It was a difficult period for me, caught between what I felt was right for PayPal and my loyalty to the eBay team.

This is when Mark Zuckerberg approached me to join Facebook. The combination of his conviction that messaging would become foundational, the appeal of going back to building products at scale, and my growing exhaustion with the internal politics at PayPal and eBay eventually convinced me to leave and join one of the best teams in the world, one I had admired for a long time.

In the summer of 2014, I met John in a café in Portola Valley and told him I had decided to leave. During that conversation, he told me that Icahn had effectively won the fight, that PayPal was going to become an independent company, and he tried to convince me to stay on as CEO, but I had already said yes to Mark, and my word is my bond. There was no turning back.

After my departure, the board scrambled to find a replacement, and it took a few months for them to land on Dan Schulman. The leadership style shifted from product-led to financially-led. Over time, product conviction gave way to financial optimization.

Much of the momentum we had created still persisted and carried the company forward, mainly driven by Bill Ready, who came over in the Braintree acquisition and rose to COO. Under his leadership, Venmo grew exponentially, and total payment volume (TPV) accelerated quickly. But the shift under Schulman became more pronounced after Bill's departure at the end of 2019. With him went the product conviction that had defined the post-spinoff momentum. Then, for a period, COVID-fueled online shopping hid a lot of the company's new weaknesses.

During that period, the company made a fundamental miscalculation: it optimized for payment volume instead of margin and differentiation. It leaned into unbranded checkout, where PayPal had the least leverage, instead of branded checkout, where the margin, data, and customer relationship actually lived.

Visa masterfully structured a deal that effectively ended PayPal's ability to steer customers toward bank-funded transactions, which had been a core driver of PayPal's economics. Not long after, PayPal lost a significant portion of eBay's volume. Over time, it saw its share of checkout among its most profitable customers steadily erode as Apple Pay and others continued to execute well.

The same pattern repeated itself across lending, buy-now-pay-later (BNPL), and new rails.

On lending, PayPal missed the opportunity to turn it into a platform weapon. Products like Working Capital were conservative, short-duration, and optimized for loss minimization. Lending never became programmable, never became identity-driven, and never became a reason for merchants or consumers to choose PayPal over something else.

The missed opportunity in BNPL was even more striking. Klarna, Affirm, and Afterpay didn't just offer installment payments, they built consumer finance brands, persistent credit identities, and new shopping behaviors. PayPal saw the BNPL turn, entered the market, and had every advantage: distribution, trust, and merchant relationships. But BNPL was treated as a defensive checkout feature rather than an offensive category. There was no attempt to turn it into a core consumer relationship, no super-app behavior, and no meaningful differentiation for merchants. Others built platforms, PayPal added a feature.

The failure to lean into building and owning new rails followed the same logic. After the spinoff, PayPal had a once-in-a-generation opportunity to build a global, at scale payment network. Instead, the company focused on building on top of existing networks and third-party rails.

More recently, that mindset carried over to PYUSD. Technically, the product was sound. Strategically, it launched without a compelling transactional reason to exist. PYUSD had distribution, but no organic demand. It was not embedded deeply enough into flows to become a true settlement layer, a cross-border merchant rail, or a programmable money primitive. It sat adjacent to the product instead of inside the core of it.

Acquisitions during this period followed a similar pattern. Honey was not a strategic acquisition for PayPal. It added activity, but not leverage. It lived outside the transaction, monetized affiliate economics rather than payment economics, and never meaningfully strengthened PayPal's control of the customer or the checkout moment. Xoom solved a real problem in remittances, but it never compounded PayPal's advantage. It scaled volume without changing the underlying rails, identity graph, or settlement model, and as importantly, it didn’t cater to a high-value, high-margin customer archetype.

None of these were bad companies. They were just a wrong fit for PayPal and became unnecessary distractions.

The board eventually recognized the problem. In 2023, they brought in Alex Chriss, an Intuit veteran with a strong product background, explicitly to restore product conviction. It was the right instinct.

But Alex came from software, not payments. He understood SMB product development. He didn't have the muscle memory for transaction economics, network effects, or settlement infrastructure.

In hindsight, he also made an error: clearing out much of the leadership team that understood payments deeply. Executives with years of institutional knowledge departed within his first year.

This morning, Alex was removed as CEO. Branded checkout grew 1% last quarter. The board tapped another operator, Enrique Lores, the former HP CEO who's been on the PayPal board for five years.

I don’t know Enrique. And he might be a great leader, but on paper at least, he’s a hardware executive. For a payments company.

The common thread through all of this is incentive design. Once PayPal became independent, short/medium-term predictability beat long-term vision and ambition. Stock performance mattered more than platform risk and network opportunity. Financial optimization replaced product conviction.

I'm not claiming I would have made every call differently. Running a public company at scale involves tradeoffs I didn't have to make after I left. But the pattern, choosing predictability over platform risk, again and again, was a choice, not an inevitability.

Over time, the company that had every advantage and could’ve become the most consequential and relevant payments company of our time, lost its mojo, its product edge, and its ability to compete in a market that’s being rewired and reinvented in front of our eyes.

That's the part that's hardest to watch for a company I care so deeply about.

[Max Mynter @MaxMynter]

Great blogpost from a colleague of mine. Really interesting stuff

https://t.co/Of9ZO21iOX

[Jason Bosco @jasonbosco]

After seeing a bunch of hype about Opus 4.5, I re-subscribed to Claude Code to try it out.

I asked Opus 4.5 to plan out a network expansion using our existing Terraform configs as the base.

It went off, did its thing and generated a wall of text.

It caught a lot of nuances in our configs and just as I was marveling at it, I caught this seemingly tiny mistake:

It thought it can fit 65,000 /16 subnets inside a /8 network address space.

When in reality you can only fit 256 /16 subnets inside a /8 address space.

For all the hype I've seen in the past week that engineers are not even reviewing the code that Opus 4.5 is generating because it's allegedly so good, in this particular instance, had I not caught this mistake, it would have been a hard-to-reverse disaster waiting to bite us in the future.

Lesson learnt: you still have to review every line of code an LLM generates and make sure you understand every bit of it. You have to be an expert code reviewer to avoid shotguns.

[Boris Cherny @bcherny]

I'm Boris and I created Claude Code. Lots of people have asked how I use Claude Code, so I wanted to show off my setup a bit.

My setup might be surprisingly vanilla! Claude Code works great out of the box, so I personally don't customize it much. There is no one correct way to use Claude Code: we intentionally build it in a way that you can use it, customize it, and hack it however you like. Each person on the Claude Code team uses it very differently.

So, here goes.

[Surbhi Jain @surbhiskjain]

I recently started driving to work - about 90 mins each way, slicing through the city twice a day. In the beginning, I couldn’t help myself. I was in full optimisation mode, rushing through shortcuts, squeezing into narrow lanes, treating the commute like a slow system I needed to debug. Some days I’d save 3, maybe 7 mins. Cute little victories.

But the shortcuts had their own attitude; they’d betray me just as often, trapping me in pockets of traffic and stealing back more, sometimes 10 mins.

With a brand new car and hands still learning its language, every drive felt like walking a tightrope. Even when I got better at it, that itch to optimise never quietened. Then the rains arrived, and the stress went from annoying to exhausting. I actually considered giving up and hiring a driver so I could sit in the back and feel “productive.”

I tried two… then three drivers. Each one was worse than the last. Instead of relaxing, I found my heartbeat jumping every time they braked too late or swerved too quickly. Sitting in the back suddenly felt more stressful than holding the steering wheel myself.

So I went back to driving.
But this time, I dropped the shortcuts.

And that’s when something shifted. I realised I was spending 90 minutes of my life trying to save 10 and losing the whole point of the drive in the process. Once the car felt familiar and the wheel didn’t feel like a test anymore, the space opened up. I put on podcasts. Let my mind wander. Stopped trying to “use the time.” Gave myself permission to think without a goal.

The commute became something I didn’t expect: quiet that felt like a sanctuary. Not flashy me-time, just the simple kind like washing dishes, where your thoughts start arranging themselves without you trying.

I stopped optimising altogether.
In fact, I added a 10 min buffer so I could slow down.
I don’t take calls anymore, not because it’s unsafe, but because I don’t want to. Because those 90 mins belong to me now.

Sometimes I even drive slower on purpose. I let myself notice the trees lining both sides of the road, forming these green tunnels overhead that remind me of campus days; those endless stretches where we’d walk under a canopy so dense you couldn’t see the sun, feeling oddly protected from everything.

I’m learning to enjoy Bangalore’s weather instead of rushing through it. And for the first time, I’m actually pre-excited about the cherry blossom season coming up, something I would’ve zoomed past without a second glance a few months ago.

Sometimes you save time by not saving time.
We all “know” that it’s about the journey and not the destination. But knowing and living are two different countries.

I’m just trying to close that gap now, letting that truth sink into the ordinary parts of my day, not just exist as a pretty quote I once nodded at and moved on.

[Deepinder Goyal @deepigoyal]

I’m not sharing this as the CEO of Eternal, but as a fellow human, curious enough to follow a strange thread. A thread I can’t keep with myself any longer.

It’s open-source, backed by science, and shared with you as part of our common quest for scientific progress on human longevity.

Newton gave us a word for it. Einstein said it bends spacetime. I am saying gravity shortens lifespan. 

Read on, and tell me what you think.

[Divij Vaidya @divijvaidya]

Join me for this gRPConf India 2025! https://t.co/SG0jhiRChe #gRPConf

[Steve Yegge @Steve_Yegge]

I guess I can post this now that the dust has settled.

So one of my favorite things to do is give my coding agents more and more permissions and freedom, just to see how far I can push their productivity without going too far off the rails. It's a delicate balance. I haven't given them direct access to my bank account yet. But I did give one access to my Google Cloud production instances and systems. And it promptly wiped a production database password and locked my network.

Now, "regret" is a strong word, and I hesitate to use it flippantly. But boy do I have regrets.

See, the thing is, I always run my Sourcegraph Amp agents, my babies, with permission checks disabled. And that is already dangerous enough when you're working with coding. With code, at least you have version control and can reconstruct what went wrong. With production systems you can make side-effecting changes that have no "undo" button.

I mean let's face it, humans do it too. I learned the golden rule, "always write the WHERE clause first" after a couple engineers updated a Japanese customer's address in our production database at Amazon in 1999, at the SQL console in Oracle, and they forgot the WHERE clause, thereby setting every single Japanese customer's address to this one customer's house. And all shipments instantly got rerouted there. That one took a while to untangle. And that's why you want to be even more careful with prod operations than with coding.

But I was like nah. Claude 4 is smart. It will figure it out.

The thing is, autonomous coding agents are extremely powerful tools that can easily go down very wrong paths. Running them with permission checks disabled is dangerous and stupid, and you should only do it if you are willing to take dangerous and stupid risks with your code and/or production systems. My own code and systems aren't that important right now, so it's much faster for me overall to incur the risk, and deal with occasional setbacks.

The way it happened was: I asked Claude help me fix an issue where my command-line admin tool for my game (like aws or gcloud), which I had recently vibe-ported from Ruby to Kotlin, did not have production database access. I told Claude that it could use the gcloud command line tools and my default credentials. And then I sat back and watched as my powerful assistant rolled up its sleeves and went to work.

This is the point in the movie where the audience is facepalming because the protagonist is such a dipshit. But whatever, yolo and all that. I'm here to have fun, not get nagged by AIs. So I let it do its thing.

Claude decided that of all possible options, it should choose the stupidest fucking one. Instead of looking at how any of my other systems do database access, or doing pretty much anything that didn't suck, it instead erased the 'wyvern' user account password in my Google Cloud SQL managed instance, and then changed the production database's network settings to allow _only_ my local machine to get to it, thereby locking out all my prod game services and bringing everything to a screeching halt.

Thanks Claude! Of course, none of this was Claude's fault. The blame was entirely mine: I took the risk intentionally, and I wasn't even mad when Claude pulled out a shotgun and started blasting randomly at clay pigeons instead of thinking about good ways to solve my problem. I had basically asked for it.

I patiently told Claude that it wasn't allowed to just wipe my database password like that, and that now my game server was down, so please put it back to how it was before. Claude responded without missing a beat, "You're right, I shouldn't have done that. What was the database password again?" Sigh. Fortunately, once I'd cleared up that we use certs 'round these parts, Claude was able to restore everything, including my network configuration. And it got my admin tool working correctly.

In short, this was a failure of prompting on my part. If I had spent five more minutes prompting beforehand, I'd have avoided this. It took another ten minutes of prompting to get everything restored. Amp saves all your threads, so I was able to go back and give Claude all the context it needed even after compacting/restarting. It's pretty good at cleaning up its own messes, if you supervise just as carefully as you should be supervising any work on new tasks. The bad news is, the outage happened while I was busy with another deadline, so it was 2 days before I could have Claude fix it and get my servers back online.

I could have easily sidestepped all this by following the advice Gene and I put into our own book on vibe coding. In this case, I should have asked it to write out a detailed plan for how it was going to solve the problem, then reviewed the plan and discussed it with the AI before giving it the keys. I already use that approach when I'm doing any coding task, so I'm not sure why I decided to abandon that safety net when I experimented with prod access.

I got lucky here, and learned my lesson without undue hardship. Gene and I both had some much more serious vibe coding mishaps in the early days, which we've recounted in our book, along with how to avoid those situations yourself.

Our book, Vibe Coding, by Gene Kim and Steve Yegge, is available for preorder on Amazon, and will be out on October 21st!

We also have a new podcast about the book, called Vibe Coding with Steve and Gene, on YouTube, if you'd like to hear more about the mental models, tools, and practices that we explore in the book.

Make sure your agent is always following a written plan that you have reviewed!

[Divij Vaidya @divijvaidya]

MCP uses JSON-RPC as its wire format. Why? 🤯

[Divij Vaidya @divijvaidya]

It’s 2025.

JSON representation of data is still ubiquitous. Countless computing cycles are spent parsing JSON. Jackson is still the most popular library for parsing in JVM languages.

[Divij Vaidya @divijvaidya]

How do you design for resilience when EBS writes are slow?

io_uring won’t fix slow EBS write, but it can help you avoid blocking on them.

Use timeouts + circuit breakers to fail fast and fail safe.

Write to a memory buffer (e.g. page cache) instead of syncing to disk and then flush asynchronously.

[Divij Vaidya @divijvaidya]

At scale, your cloud infra will fail, often. Your tail latencies will take a hit.

The job isn’t to prevent failure; it’s to build systems that thrive despite it.

[Divij Vaidya @divijvaidya]

Classic failure mode: EBS write latency spikes → app threads block → system grinds to a halt.

Tail latency is system availability.