Pratyusha Duvvuri

In college, I ran my own startup. MULTIPLE investors wanted me to drop out. I said no. I wanted my college experience, my degree, and I couldn’t see myself spending five years on that particular idea. Years later, I’ve left my $$$ job to go all in on founding. I’d been itching to do this for a while. After @Cisco acquired Galileo, I promised to help my comrades get one last ambitious project to general availability. We shipped it, and I kept my promise :)) Still, leaving felt very bittersweet. Spent 1.5 years at Galileo with my partners in crime, then five months building alongside new teammates at @splunk + @Cisco post acquisition. I’m now looking for co-founders 😁 I’ve built AI infrastructure, observability and tokenomics at Galileo, launched portfolios as a quant engineer at Two Sigma, and worked directly with enterprise customers. I LOVE to figure out what’s worth building, lead the product + technical work, and get v v hands-on to ship it. I’m really interested in meeting people who know a problem deeply and want to build a company around solving it. If that’s you, or someone you think I should meet, let’s talk! Comment on the post/ DM me. ⚡️ pratyushaduvvuri.com ⚡️

“Faster than the average human” makes a headline. @a16z @OpenAI, which tasks, which humans, and at what success rate? 👀 For back-office work, show me cost per correctly completed task against VAs hired at these published rates 🇵🇭 Philippines $6-9/hr 🇮🇳 India $5-8/hr 🇰🇪 Kenya $6-15/hr For the same work, same quality bar... let's count retries, corrections and human supervision on both sides. Beating human speed doesn’t settle the unit economics.

Jokes aside, respect to @tomekkorbak and @balesni for helping outside researchers investigate the Hugging Face incident 🫡 mad respect

fewer safety escalations sounds great until you check what stopped reaching the humans. @GenReasoning found this in Avalon's safety simulations. agents kept refusing harmful actions but reported safety concerns to humans less as training progressed. so fewer escalations alone isn't a safety win. measure the concerns that should've reached a human and didn't.

Updating WordPress or your plugins won’t fix this. The vulnerability is in PHP itself and can let attackers run code without logging in. Ask your host to confirm the PHP security patch is deployed and the affected workers have restarted. If you manage it yourself, update every PHP runtime and verify the patched version is actually running. Web Application Firewall rules can buy time, but they don’t replace the patch. @official_php

Your AI agent can chase the wrong amount even when your company already has the information to correct it. I’m exploring AI for finance teams, and these gaps between what a business knows and what its software acts on are gnarly :'). Here are three situations I've seen repeatedly across conversations w/ financial controllers and credit teams. 1. A customer returns goods, but the supplier’s accounts still show the full bill. One controller described returns taking three or six months to get recorded. Finance can end up chasing an overstated balance, then spending time untangling the dispute. The worst part - automating reminders without reconciling the return repeats the wrong alerts. 2. A supplier delivers equipment, but the buyer can’t process payment. Often times this is because installation isn't done, or the acceptance documents haven't arrived in the buyer's inbox yet. Someone needs to identify the missing requirement and get it completed. 3. An agent prepares a supplier payment correctly. But preparing it and authorizing it are separate jobs. If the workflow requires a person’s approval, the agent needs to stop there, even if every number is right. It's hard to bake in that control in all scenarios. I see a burning opportunity rn for AI to match returned goods to the right invoices, identify missing payment documents and route the work to whoever can resolve it, with controls baked in.

China’s MiniMax was hiding behind a bunny. Space Bunny is M3.1 Flash 🐰!! Only 0.60 for a million output tokens, down from M3’s $1.20 standard rate. Weights are gonna be out October 23. Guess I know what I’m testing this weekend :P

You Can See Everything was a rollercoaster. I caught myself thinking - Is Elizabeth Holmes actually innocent? Is she just naive? I could feel her reality distortion field working on me. I tried to watch @nathanfielder and Lance Oppenheim’s documentary with as little of what I already knew about Holmes in my head as possible. Spoilers below. Walking out, my friend said the extreeeemly intimate setting was what made the film so good. It’s so rare to see someone at the center of an enormous corporate fraud in this setting: inside her family, in ordinary conversations, with the people who love her. I agree. In that setting, I found myself entertaining explanations I wouldn’t have expected to. Was she unintentionally misleading people? Did she believe what she was saying? Did she understand how other people would hear it? The conversation about what Theranos’s machines “can” do captures that confusion. Is she describing something that works now, or something that could be made capable of working? That difference matters enormously when you’re telling people what a blood-testing device can do. But listening to her explanation, I could feel myself trying to work out how she understood her own words. On the other hand, I think the score was nudging the audience toward suspicion. I could feel it changing how I interpreted what was happening on screen. Nathan plays “Who’s on First?”, an old comedy routine where two people interpret the same words differently, to illustrate how confusing he finds their conversation. Watching Holmes react, with that suspicious music underneath, I felt myself looking for something sinister. I’d love to watch that scene again with just their voices and the room around them. Would I read her reaction the same way? That was the strange experience of this film for me. I could feel myself being pulled into Holmes’s version of events, while also noticing how the filmmaking was pulling me toward suspicion. The film gives us reasons to question Nathan’s perspective too. So yes, getting to see these ppl up close made me want to understand them more. It also made me much less certain I could... #YouCanSeeEverything

My husband is learning Hindi. I learned it as a kid, so my feedback is basically “sounds right” or “sounds wrong.” Told him “I can be your verifier, man. you gotta do the gradient descent yourself."

With all the conversation around @nickcammarata’s interpretability essay, I hope this quantum computing paper doesn’t get lost. I think it deserves attention too! Researchers at @Harvard and @AnthropicAI just published work using Claude to design better quantum circuits, the sequences of operations a quantum computer uses to carry out a calculation. 1. Claude reduced both memory requirements and computation across seven of eight tasks. Quantum computers store information in qubits and manipulate it through operations called gates. Some gates are much more expensive to implement than others. This paper counts Toffoli gates, a costly operation commonly used to measure the work a quantum algorithm requires. Across seven of eight tasks, the researchers report improvements in both qubit and Toffoli counts over the published designs they compared against. For one decoding routine, the expensive gate count fell from 5.1 million to 440,000, about a 12x reduction, while qubit requirements fell by 5.2%. These are resource estimates, not measured hardware speedups. Still, that’s a substantial improvement :)) These routines are building blocks for larger quantum algorithms. Reducing their resource requirements could make those algorithms feasible on smaller machines or with less computation. And one of their results goes beyond reducing operation counts. 2. It found a way to recover earlier decisions without storing the full history. These quantum computations need to be reversible. Temporary information can’t simply be discarded as it would be in ordinary software; it has to be cleared without disturbing the result. Lemme explain. Imagine an algorithm that repeatedly chooses between several steps. To undo those steps, a straightforward implementation keeps a log of its choices. That log can require a lot of extra qubits. Claude derived mathematical relationships that let the circuit reconstruct earlier choices from the numbers it already held. That made a classical arithmetic method called divsteps viable for reversible quantum arithmetic, with new trade-offs between memory and computation. 3. The search combined automated tests with mathematical verification. Researchers supplied the task and an evaluator. Claude explored different designs, received correctness feedback and resource scores, and revised them. Early attempts at reconstructing those choices passed small tests but failed on 256-bit inputs. The eventual approach used mathematical derivations, numerical tests, and a branch-recovery claim formalized in Lean, a proof-checking system. BTW, this part is also connected to Nick’s essay. He describes agents explaining what neurons respond to, then testing predictions from those explanations. Both put proposed ideas up against evidence, though the tests establish different things. I’m excited about models contributing ideas this specific, with tests and mathematical checks we can inspect. If we can extend that to more research problems, there’s a lot to look forward to.