Rishi Kulkarni

You should be changing your mind about AI at least every 12 months. zou-group.github.io/llm-clai…

Today’s @timesofindia is a shameless collection of advertising- first3 pages & then news on left page & ads on right on all pages, incldng the one facing op-ed page. Is this a newspaper any longer or news used to market ads. PATHETIC !

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…

In the year 2000, Joel Spolsky (founder of Stack Overflow) wrote a famous article titled “Things You Should Never Do, Part I.” In it, he talks about how rewriting your code from scratch is the single worst strategic mistake any software company can make. However, times have changed. The cost of writing software is dropping exponentially and that has changed the math entirely on if it is worth rewriting your codebase. For instance, this year Anthropic’s Jarred Sumner ( @jarredsumner ) rewrote Bun in Rust in 11 days leading to many performance and correctness benefits. Likewise, Stephen Toub (Distinguished Engineer at Microsoft, @stephentoub ) rewrote the GitHub Copilot runtime (400,000-line codebase) in Rust by himself. I thought it’d be interesting to lay out everything he learned in the process. So I interviewed him about how and why he did it. In this episode: • Why he choose Rust of all languages • Rollout strategies (incremental vs all at once) • How he prevented regressions • Steering agents and which models he used • Career lessons and engineering advice Where to watch: • YouTube - youtube.com/watch?v=PLMOVlZC… • Spotify - open.spotify.com/episode/4zt… • Apple Podcasts - podcasts.apple.com/us/podcas… • Transcript - developing.dev/p/e073-stephe… Thank you to the sponsor of this episode for supporting my work: • WorkOS: makes your app Enterprise Ready with easy to use APIs to add SSO, SCIM, RBAC, and more in just a few lines of code, check them out at workos.com/ Chapters: 00:00 Intro 01:06 Why the Copilot runtime needed a Rust port 05:38 Why he chose Rust 11:28 Discussing rollout strategies 21:57 How he caught regressions 32:08 Guiding the agents 42:11 Catching regressions before prod 45:16 Which models he used 50:03 The end benefit and costs 01:03:24 When to do a full rewrite 01:06:40 Career learnings on the way to Distinguished 01:16:00 Top technical book recommendation 01:17:49 Advice for young engineers 01:21:00 Outro

At @temple, one of our missions is to become the ultimate wearable for elite performance athletes. And today, let's start with the basics: Heart Rate (HR). You would think accurate HR tracking is a solved problem. It is not. Especially in zones 4 & 5, which matter to professional athletes the most. To see whether Temple is the most accurate wearable on something as basic as HR, we gave ourselves a high bar to aspire to. A Polar H10 chest strap, taped to the chest. Taped, because worn the usual way, straps lose contact mid-sprint and miss beats. We benchmarked five flagship wearables against that reference, from rest to max effort, second by second. All the other leading wearables were well behind Temple in our study. Apple Watch Series 12 came closest. Temple also beat the untaped chest strap, the industry's benchmark, in zones 4 and 5. Why does Temple win? Two reasons. 1. Temple uses adhesive, so its lab condition is its wear condition. You'll experience our study outcomes in real life. 2. The location. Your temple sits right over a major artery, and it's the one patch of skin that doesn't clamp down its blood flow when you go hard. We tested the same sensor at four body sites. The temple was the most accurate of the four. Charts attached. Accurate HR makes Temple better at so much more downstream. More soon. The full paper and data are live right now at temple.com/science, and an extended study by independent researchers is underway. We invite further independent replication in more conditions, and will fund it. Write to us at [email protected]

After HCLTech led Sarvam’s $300m round in June at a $1.5b valuation, Cognizant is reportedly in talks to join the next one, which would value Sarvam at $3.5 to 4 billion, more than double, with large Indian enterprises lined up to anchor it. For an industry trading near five-year lows on fears that AI agents will automate away coding, maintenance and back-office work, good to see them backing homegrown model builder is both a hedge and a bet on the future. via @EconomicTimes

We've reached a point where people who've never built a company have absolute certainty about how companies should be built. Meanwhile, founders who've spent years building one will openly admit how much they still don't know. The internet has created a strange world where those with the least experience often have the strongest opinions, while those with the most experience are expected to defend every word they say. It's remarkable how much easier entrepreneurship looks when the only thing you've ever built is an opinion about it.

every product i use should just be a headless mcp, that feeds into an agent, where i never have to think about anything underneath. except for our product, of course. in our product, customers love the ui and love us very much. we're special that way.

This is why I recommend younger founders to build profitable, bootstrapped business, make money, and then think of building a moonshot, venture-scale business. I did exactly this myself. Built my first venture, sold it, made good money and became financially free. Now I'm building a moonshot at @GreyLabsAI taking a big goal of disrupting the call center industry using AI. Worst case if it doesn't work, I won't find myself on the road. Best case if it works, it's going to be a great outcome for all.

Exa now indexes 3,000,000+ podcasts. Your agents can now search across millions of hours of transcripts, updated daily, and track what the world is talking about.

4 months ago I cloned Airtable for a client in 5 days because they refused to pay Airtable $500/mo months later, I’m still maintaining it every two weeks it’s another edge case, a UI tweak, or a "quick 10-minute request." They’ve easily spent 10x more in headaches than the $2,000 they would’ve paid Airtable just proving that SaaS isn't dead. Cloning isn't enough, if you want it to be same quality as original it becomes a full time job and at this point isn't it easier to just pay?

The exploits of the last few days are a good reminder of why we use SaaS. Our company pays for ~20 SaaS services. If we self-hosted them, we’d now be responsible for all of the security patches this week. Instead, our small monthly bill pays for their teams to handle everything.

Sadly canceling our linear subscription as we just don’t use it anymore Planning happens in a Slack canvas with very high level goals and constraints Features gets shipped in a few hours, or days max Roadmaps don’t get further than a month

Reading a history of the bond market recently, and the author noted that Britain defeated Napoleon mostly because their consol bond market left the British military much better funded than France’s weak post-revolutionary capital markets. One Alexander Hamilton studied the British bond market, and then consolidated American revolutionary debt into one tradable market, which gave birth to what became the New York stock exchange, Liberty bonds, and the world’s largest capital markets. An army might march on its stomach, but it’s borrowed money that pays for the food and ammunition, and the one who can borrow the most generally wins.

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scratch in under 4 months Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

Maturity arrives as soon as we can laugh, rather than fret, over our insufficiencies, or those of others. Growing up means throwing one's arms around the imperfection of almost everything.

Everyday I am reminded how much damage was done to generations of Indian youth because the smart people were told to be quiet and humble and let their work speak, while the idiots were given free rein to have a personality.

These days, I'm struck by how the dominant mood on both sides of the Atlantic toward the other is a kind of pity. Europe frets about the US's supposed slide into autocracy; the US about Europe's pusillanimity and immiseration.

You already gave the LLM the answer. So why is it writing it out from scratch? Introducing a new AI model architecture: Constrained Decoding Models (CDMs) LLMs were built to write. Most production workloads just need an answer: fast, consistent, and affordable. CDMs are purpose-built for 1 task each. They read your input once and skip the token-by-token generative loop that makes LLMs slow and expensive. Not just a decision model, but a family of 3 CDMs live today: • Classify: define your labels and get back a calibrated probability + reason for each. Used for routing, tagging, and intent detection. • Extract: describe the entities in natural language and get back well-formed JSON with confidence scores + citations. Used for web scraping, document processing, and PII redaction. • Summarize: pass in long text, images, or PDFs and get back a consistent summary under 1,000 tokens. Used for context compaction, metadata generation, and transcript digests. Built for production: • 1M tokens of multimodal (text + image) input, 31x larger than Jev • 77.2% on DecisionBench vs Jev's 72% • 200x faster, 467x cheaper than frontier LLM APIs • $0.05/M input tokens. Output is free (yes, even for summarize) Live on your favorite platforms today: @OpenRouter @vercel_dev @LiteLLM @LangChain @TrustedRouter @itsArthurAI @useRouterPlus Get 50M input tokens for free (for the first 50 hours) at link below 👀

Technical moats are mostly over at this point. If you’re building something that is meant to be durable, the only moats now are: • Proprietary data • Network effects • A physical component