Advait Maybhate

👾 𝗥𝗲𝗰𝗲𝗻𝘁𝗹𝘆, 𝗖𝗨𝗥𝗟 𝘀𝗵𝘂𝘁 𝗱𝗼𝘄𝗻 𝗶𝘁𝘀 𝗯𝘂𝗴 𝗯𝗼𝘂𝗻𝘁𝘆 𝗽𝗿𝗼𝗴𝗿𝗮𝗺 𝗮𝗳𝘁𝗲𝗿 𝗴𝗲𝘁𝘁𝗶𝗻𝗴 𝗳𝗹𝗼𝗼𝗱𝗲𝗱 𝘄𝗶𝘁𝗵 𝗔𝗜 𝘀𝗹𝗼𝗽 𝗿𝗲𝗽𝗼𝗿𝘁𝘀. I spoke with CURL’s founder about what this means for the future. Daniel Stenberg recently wrote a blog explaining why CURL needed to shut down the program due to AI slop reporting. Naturally, it sparked a big debate about where bug bounty goes from here. CURL is one of the foundational pieces of internet plumbing, started as a personal project in the 90's, today 𝟯𝟬 𝗯𝗶𝗹𝗹𝗶𝗼𝗻 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀 𝗿𝗲𝗹𝘆 𝗼𝗻 𝗶𝘁. CURL’s bug bounty program, since 2019, played a key role in securing the project, and 𝗗𝗮𝗻𝗶𝗲𝗹 𝘀𝗮𝘆𝘀 𝗵𝗲 𝗮𝗹𝘄𝗮𝘆𝘀 𝗰𝗼𝗻𝘀𝗶𝗱𝗲𝗿𝗲𝗱 𝗶𝘁 𝗮 𝘀𝘂𝗰𝗰𝗲𝘀𝘀. So why shut it down? In this episode we discuss: • Why curl shut down bug bounty payouts after a surge of AI-generated reports • How convincing AI reports are actually **harder** to triage than obvious bad ones • What this might mean for the future of bug bounty programs • How curl started as a tiny tool for an IRC bot… and ended up running in billions of applications • Whether Daniel ever regrets starting the project that became his biggest legacy A great conversation about open source, security, and how the internet’s plumbing actually gets built. #BugBounty #OpenSourceSecurity #AppSec #AIinSecurity #curl

141

7 Comments

Data is what’s holding back AI for climate. At PNW Climate Week I heard this again and again. Climate Change AI just published a community-curated list of the data gaps, from power grids optimization and climate policy records to marine species identification. Check if your field is covered, and if not, contribute to it. Link to page: https://lnkd.in/gBGyGagq #ClimateChangeAI #Data

6

Most people think a “fingerprint” is just metadata you can delete. It’s not. Modern systems turn images into high dimensional math vectors. You’re not fighting files. You’re fighting linear algebra. That’s why reverse image search works. That’s why AI detection works. I still strip EXIF data. That removes obvious traces like location and software. But the real signature? It lives in the math. Most people don’t even understand what a fingerprint actually is. Do you?

1

Today we launched Liminary to open beta. When you’re researching a new topic, how much time do you spend on “paperwork?” Hunting for that smart blog post you saw last month, cross-checking sources for areas of consensus and dispute, making sure your quotes are accurate, etc. With Liminary, spend less time on paperwork and more time drawing conclusions. - Save sources and notes. - Ask questions, get summaries, compare perspectives. - Recall items from your library as you browse for new sources or draft a report. Try it out at liminary.io. Let me know what you think. (If you’re on mobile, use a standalone browser app to sign up.)

13

2 Comments

Recently I gave a talk at SF Ruby about a library we built (and recently open-sourced!) at AngelList called Zaxcel. Why'd we build it? Excel is the lingua franca of finance, but most programmatic Excel generation libraries force you to think in coordinates and formula strings. We took a different approach: model the workbook as a directed graph where data flows through named objects. Let the library handle coordinate resolution and formula composition. The result is a DSL that actually makes sense: you write formulas as Ruby expressions, reference things by name, and the type checker catches mistakes at compile time. It's revolutionized how we build financial statements. What used to take weeks now takes hours (and with AI agents - only minutes!), empowering thousands of workbook deliveries a quarter. Read how we went from nightmare refactors to 100x productivity: https://lnkd.in/es2aTttR Open-sourced repo (you can use it today!): https://lnkd.in/ecKRwcaW

90

8 Comments

👉 LLMs don’t feel mysterious once you break them down, they’re built by scaling a handful of simple ideas. Taking the next step in learning how language models are built from scratch, I implemented a character-level MLP. This step was where things really started to click. Moving beyond the bigram model, I built a character-level language model using a simple MLP, learning token embeddings, passing them through hidden layers, and predicting the next character via a SoftMax output. What stood out immediately was the difference in learning capacity. By introducing an MLP: 1. The model could capture richer context 2. Embeddings became learned representations, not just counts 3. The training loss dropped significantly compared to the bigram baseline Seeing the loss curve degrade (in a good way 😄) much faster than the bigram model was a clear reminder of how even a small neural network can dramatically improve expressiveness. This part also reinforced key ideas: 1. Why embeddings matter 2. How non-linearities unlock better representations 3. And how gradient descent shapes both embeddings and network weights together Building this from scratch made it clear that modern language models are still rooted in these exact ideas, just scaled up massively. Notebook : https://lnkd.in/df7yCEas Next up: pushing Makemore further and understanding how these foundations evolve toward deeper and more powerful models 🚀 #DeepLearning #MachineLearning #NeuralNetworks #Backpropagation #Autograd #Micrograd #MLP #AIEngineering #LearningInPublic #BuildInPublic #FromFirstPrinciples #AndrejKarpathy #GenAI #AICommunity

27

Folks, check this out! AI for video is huge and we WILL one day tell 95% plus of stories via content co-created with AI.

5

Curvture Variable Physics - CVP is a unifying control framework—toroidal refractance and resonance scheduling applied consistently from bio-nanonetwork signaling to quantum networking and fault-tolerant computation. Our filings are designed to protect the core architectures and methods required to translate this framework into deployable systems.”

2

1 Comment

Musician Rick Beato explain how music studios, that took millions to build and that charged musician tons of money to use it run out of business when home studios arrived. His take is that AI datacenters that are being built now already obsolete with LLMs that may run locally. Interesting perspective. https://lnkd.in/esScgRe4

2

We’ve all been there, 𝘀𝗲𝗻𝗱𝗶𝗻𝗴 𝗵𝘂𝗴𝗲 𝗝𝗦𝗢𝗡 𝗯𝗹𝗼𝗯𝘀 𝘁𝗼 𝗮𝗻 𝗟𝗟𝗠 and wondering why the prompt feels so... heavy. I recently came across something that caught my eye: 𝗧𝗢𝗢𝗡 (𝗧𝗼𝗸𝗲𝗻-𝗢𝗿𝗶𝗲𝗻𝘁𝗲𝗱 𝗢𝗯𝗷𝗲𝗰𝘁 𝗡𝗼𝘁𝗮𝘁𝗶𝗼𝗻), a new, lightweight data format designed specifically for LLMs. Unlike JSON, TOON doesn’t waste space repeating keys or wrapping every value in braces. It keeps the structure, but strips away the fluff, kind of like YAML meets CSV, tuned for how LLMs actually read data. It’s clean, readable, and the models parse it with fewer errors - especially when generating structured outputs. For devs working with large or repeated data, this could be a game-changer. 𝗧𝗢𝗢𝗡 𝗶𝘀𝗻’𝘁 𝗺𝗮𝗴𝗶𝗰, if your data is full of nested objects, it might actually take more tokens than JSON. But it really shines with flat, consistent data. 💭 Will LLM-friendly formats like TOON replace JSON, or will teams stick with the familiar? #NextGenAI #SmarterTOKENS #AIInfrastructure #SmartData #FutureOfCoding #TOON

43

One of our early beta users shared something with me that stuck: He’s been using metra.ai weekly to monitor his startup team He’s looking for signals: - Who’s doing well - Where work might be slowing down - What needs his attention before it turns into a fire That’s exactly what we built this for. Not to track time. But to give managers real visibility,without relying on gut instinct or constant check-ins. Early stage teams move fast. Metra helps them stay ahead of what’s coming.

5

Just spent the last few weeks deep in the voice AI trenches, and I keep having the same conversation with other engineers. We're all tired of the black box. That's why I've fallen hard for Pipecat. It’s not just another platform—it’s an open-source framework that gives you back the keys to the kingdom. You get to build what you want, how you want it, without your hands tied. I put together a 10-page PDF breaking down exactly why. Inside, you'll find: ▸ The real voice agent lifecycle, from audio in to audio out. ▸ How Pipecat's frame-based architecture actually works (it's genius). ▸ My step-by-step guide to deploying on Pipecat Cloud. ▸ The honest, no-BS cost comparison with Vapi and Retell. ▸ The one scaling setting that kills cold starts. (Spoiler: min_agents is your best friend.) For engineers, the choice is becoming clear. Do you want a pre-fab house, or do you want to architect your own? I know which one gets me excited to code. What about you? Are you feeling the pull toward open-source, or is the convenience of managed platforms still winning? Grab the PDF below and let me know where you land. shoutout to Aleix Conchillo Flaqué and Pipecat to help develop voice agent. #VoiceAI #AIAgents #Pipecat #OpenSource #VoiceEngineer #DeveloperTools #Vapi #Retell #AI

16

Discover how to use Airtable’s brand-new AI Labs Web Table with Web Data feature to automatically search the internet and build structured tables. In this video, I'll walk through enabling the AI Labs settings, writing an optimized prompt, and letting Airtable AI populate a base with real funding and grant opportunities for ranchers and farmers. You’ll learn: ✅ How to turn on AI Labs in your Airtable workspace ✅ How to create prompts that return structured, usable data ✅ How Airtable automatically searches the web and builds a table for you ✅ A real example: finding funding and grants for farmers and ranchers Whether you’re managing agricultural projects or just curious about the future of Airtable AI, this tutorial will show you what’s possible when you combine AI + Airtable for powerful web-to-table automation. #Airtable #AirtableAI #NoCode #Automation https://lnkd.in/gWfJtqbH

3

1 Comment

Just wrapped up an exciting learning project exploring California's housing market through data science! I'm thrilled to share that my machine learning research on California house price prediction has been published by Data And Beyond! Spending weeks diving into this project with 20,000+ California housing records, I built models that revealed fascinating market insights. What started as a simple weekend experiment turned into a comprehensive study that Data And Beyond selected for publication. The results were eye-opening: median income accounts for nearly 48% of price variation, while proximity to the coast creates dramatic price differences that can be 3-5x higher than inland areas. My Random Forest model achieved 81% R-squared, while the linear regression model showed 92% prediction accuracy with only 7.8% error when tested on real data. Being recognized by Data And Beyond feels like meaningful validation of the work's quality and methodology. Beyond the technical achievements, this project really highlighted how data science can reveal patterns hiding in plain sight about one of America's most expensive housing markets. Whether you're curious about real estate trends, exploring machine learning, or interested in seeing what stories data can tell, there's something fascinating about turning raw numbers into actionable insights. Read the full study: https://lnkd.in/gwWqU-Yh #DataScience #MachineLearning #Python #RealEstate #TechProjects #ResearchPublication #PredictiveAnalytics #CaliforniaHousing #StatisticalModeling

9

3 Comments

I've been building a multi-agent research system lately and wanted to share what went into it. The interesting problems were around giving each agent its own clean context so they don't pollute each other, pausing a graph mid-execution to wait for human input before continuing, and actually being able to see what's happening inside the system when something doesn't behave as expected. Ended up using LangGraph for orchestration, Qwen3-30B running locally through LM Studio so no API costs at all, and Langfuse for observability which honestly turned out to be the most eye-opening part of the whole thing. Wrote everything up with the architecture, the code, and screenshots of the traces. Link below if you're curious 👇 https://lnkd.in/dQQgc96S #AIEngineering #LLM #MultiAgent #LangGraph #OpenSourceAI

40

4 Comments