Jonathan Levi

High-level Architecture of Claude's Advanced Research System Ever wondered how Claude can generate in-depth, high-quality reports in minutes? This diagram breaks down how multi-agent collaboration works under the hood using Claude’s advanced architecture. Here's what happens: -> User Request: A complex prompt is submitted (like listing the top 50 AI tools in education for 2025 with detailed attributes). -> Multi-Agent System Kicks In: The Lead Agent (orchestrator) breaks down the task, manages memory, tools, and assigns tasks to subagents. -> Multiple Search Subagents operate in parallel, digging up relevant data across the web. -> A Citations Subagent ensures every fact has a credible source. -> Memory allows it to track context, avoid duplication, and refine output. -> Final Output: A polished, structured, and source-backed report is returned to the user - all autonomously. Follow Nikhil Kassetty for more tech updates ! #Claude #AIArchitecture #MultiAgentSystems #FutureOfAI #NLP #PromptEngineering #ArtificialIntelligence #AdvancedResearch

124

19 Comments

Everyone noticed when NVIDIA set a record of losing most market cap in a day ($600B). This was prompted by DeepSeek AI release of LLM models which took $5.6 mm to train as opposed to hundreds of millions that current reigning champions took. In addition, DeepSeek has announced pricing which is about 1/30th of OpenAI for its use. Two things to consider 1. How did DeepSeek achieve this feat? 2. What is the impact of this achievement? 1. DeepSeek has put detailed description of its approach https://lnkd.in/e_YXVnac. And also open-sourced its code. The key elements seem to be -- Use of 8 bits (FP8) instead of 32 -- Use of mixture of experts (MoE) instead of a dense architecture (basically smaller sub models, instead of one model for everything) -- Engineering of GPU usage, balancing compute and communication overheads -- Training using reinforcement learning (without supervised human data) and then refining using supervised data -- Prediction of multiple tokens at a time Many of these ideas have been tested in US companies for a few years. But none had put them together along with RL use at scale before 2. What is the impact? NVidia lost value as people immediately assumed that demand for chips will not be as robust if the training can be done with much less compute power. It is unclear--most of the compute is used for answering questions, not in training. Also going forward, it is unclear, how broadly will these capabilities be used. One can imagine these models replacing all knowledge work, in which case the demand for doing that work is gigantic, and will continue to to grow. This also shapes the research direction and proves that "general" one size fits all, may not be the best approach. Mutliple smaller expert models might be the way to go forward. This is great for innovation. At Estee Advisors we have been working to build a trading bot, and a financial advise bot. Happy to hear your thoughts on DeepSeek's achievement and its impact.

47

2 Comments

At BlueFlame AI, we frequently get asked about real-world applications of different LLM models. Today, I want to share a fascinating test we conducted on a complex venture capital calculation scenario. The Challenge: We presented various LLMs with a detailed Series A to Series B anti-dilution calculation problem - the kind that typically requires 20+ minutes in Excel and careful verification. The task involved calculating ownership percentages, pre/post-money valuations, and anti-dilution rights for a $10M Series A round (at $100M post-money) moving into a $50M Series B round (at $500M valuation). Models Tested: 🔹GPT-4o 🔹Anthropic's Claude Sonnet 3.5 🔹GPT-o1 and o1 mini 🔹o3 Mini 🔹DeepSeek R1 🔹Amazon's Nova Pro 🔹Google's Gemini 2.0 Flash and 1.5 Pro (v2) 🔹Meta's Llama 3.3 (70b) 📊 Key Findings: Only three models nailed the complex calculations: ✔️O1 ✔️O3 Mini ✔️DeepSeek R1 Performance Evolution: Just one month ago, only O1 could handle this type of calculation. Now we have multiple viable options showing the rapid advancement in reasoning capabilities. O1 had the most comprehensive output and cleanest tables, while DeepSeek got it right fastest (almost 5x faster than O1) 💡 What This Means: The landscape of AI capabilities is evolving rapidly. We're seeing impressive accuracy in complex financial calculations - in some cases surpassing traditional methods. The emergence of new, powerful reasoning models is particularly exciting for the future of automated financial analysis. Future Applications: Imagine automating investor documentation, calculating buy-in rights, and generating legal forms - all with precision and speed. We're getting closer to this reality. We'll continue sharing our findings as we test different use cases across various models. Each has its strengths, and we're committed to helping you navigate the optimal solutions for your specific needs. #ArtificialIntelligence #FinTech #VentureCapital #Innovation #TechnologyTrends #AITesting

69

Kronos: A Foundation Model for the Language of Financial Markets Pretty cool an open-source foundational model for financial candlesticks that was trained on data from 45 global exchanges. If you're into quant trading, this should definitely be interesting to you 😉 🔤 Github: https://lnkd.in/gcgxifg4

29

8 Comments

While others are pitching AI startups, Google just shipped an AI empire - at scale! 13+ real products. Zero vaporware. And an execution scale that only #BigTech with distribution muscle can pull off. Here is what came out at the #GoogleIO2025: ▶️ AI Mode in Search: This isn’t “search” anymore. It is becoming your digital assistant. - Context-aware. - Conversational. - Actionable: book tickets, shop, and plan your day. Already rolling out in the US, used by 1.5 Bn monthly. ▶️ Gemini 2.5 (Pro + Flash) Two variants. One goal: performance without hallucinations. - Flash: Faster, leaner, yet still brilliant on code, multimodal inputs, and reasoning. - Pro + DeepThink: Designed for deeper logic and longer tasks. Already inside Chrome, Search, Workspace. ▶️ Agent Mode (Gemini App) - Not just a chatbot. - It reads your tasks, surfs the web, books reservations, and collaborates across tabs. Think of it as your digital analyst, ops exec, and EA - all rolled into one. ▶️ Project Mariner - Now Multi-tasking - #AI that doesn’t just browse. - It watches, learns, clicks, fills forms, and acts. New update: handles multiple tasks across tabs. Only available to Google AI Ultra users. ▶️ Veo 3 - AI for Video SOTA (State-of-the-Art) video generation. - Native audio. - Camera movement. - Character continuity. Type → Get short film with audio. ▶️ Flow - Filmmaker’s Dream - Combines the best of Veo, Imagen, and Gemini. - You get a script → direction → audio pipeline in minutes. Built for creators, but usable by anyone with a story to tell. ▶️ Imagen 4 - Next-Gen Visuals - Ultra-sharp hands, lighting, text, and realism. - Better than any model we’ve seen in open-source or private benchmarks. ▶️ Lyria - Text-to-Music Just describe a mood - it creates full tracks: vocals, instruments, mixing. - Used by real artists already. - Now that’s what generative should mean. ▶️ Android XR Smart Glasses - Translate in real time. Navigate your route. Remember what you saw last week. Context-aware AR is no longer sci-fi. HP + Gentle Monster partnerships dropping soon. ▶️ Google Meet + Real-Time Speech Translation - Match tone, rhythm, and language. - Yes, it adapts your emotional delivery, too. Rolling out to AI Pro & Ultra subscribers now. Pure magic! ▶️ Smart Replies in Gmail (That sounds like you) - Gemini pulls your context from across apps to reply in your tone. - Finally, the AI doesn’t sound like AI. ▶️ Google Beam - 3D Video from 2D Calls - 2D → Real-time 3D with accurate head tracking. - Creates lifelike video calls with spatial presence. Think Zoom call. But you feel like you’re in the same room. Google is not selling an #AI app. They’re integrating AI into your daily rituals - search, docs, email, browser, phone, glasses. This is what happens when you embed AI across billions of user flows. And here’s my final take: "If you’re building an AI startup, you better assume #Google is now your competitor." #GoogleIO2025 #AI #Gemini #Google

50

3 Comments

Why Rust DOMINATES Market Making: ✅ Zero-cost abstractions (speed of C, safety of Python) ✅ No garbage collection (ZERO pause times) ✅ Memory safety without overhead ✅ Fearless concurrency (race conditions impossible) ✅ 15MB binary handles 100k orders/sec Real Performance Numbers: Order placement latency: 0.8ms (Go: 5ms) Memory usage: 50MB for 1M orders (Node: 2GB) CPU usage: 3% maintaining 10k orders Crashes in production: 0 (Zero. Nada. None.) My Rust MM Stack: rust use tokio_tungstenite for WebSockets use rust_decimal for precise math use dashmap for concurrent orderbooks use rayon for parallel processing use serde for blazing fast JSON The Secret Sauce Code: rust #[derive(Clone)] struct MarketMaker { orders: Arc<DashMap<OrderId, Order>>, spread_bps: Decimal, } impl MarketMaker { async fn quote(&self, mid_price: Decimal) -> (Order, Order) { let spread = mid_price * self.spread_bps / dec!(10000); let bid = Order { price: mid_price - spread/dec!(2), size: self.calculate_size(&mid_price), side: Side::Buy, }; let ask = Order { price: mid_price + spread/dec!(2), size: bid.size, side: Side::Sell, }; (bid, ask) } } Why Market Makers Choose Rust: 🎯 Predictable latency (no GC surprises) 💎 Rock-solid reliability (if it compiles, it works) ⚡ Optimal memory layout (cache-friendly) 🔒 Thread safety guaranteed at compile time 📊 Zero-overhead async/await Client Results After Switching to Rust: Profitability: +34% (better fill rates) Infrastructure costs: -60% (smaller servers) Downtime: 12 hours → 15 minutes/year Order throughput: 10x increase The Uncomfortable Truth: Your Python bot making markets = amateur hour Your Rust bot making markets = institutional grade Rust forces you to handle every edge case. In market making, that's not a bug—it's a feature. The compiler is your risk manager. Memory safety is your insurance. Zero downtime is your edge. #rust #rustlang #crypto #marketmaking #algorithmictrading #trading #cryptocurrency #bitcoin #ethereum #defi #blockchain #quanttrading #algotrading #fintech #tradingbot #web3 #programming #developer #tech #software #usa #uk #canada #singapore #dubai #india #australia #germany #switzerland #japan #hongkong #netherlands #france #brazil #mexico #southkorea #china #uae #binance #coinbase #kraken #dydx #hft #golang #python #javascript #cpp #systemsprogramming #lowlatency #technicaltrading #liquidityprovider #quant #hedgefund #investment #daytrading #automation #passiveincome #financialfreedom #wealthbuilding #sidehustle #remotework #techstartup #100daysofcode #buildinpublic #opensourcecode #github #developerlife #coderlife #softwareengineering #performanceoptimization #concurrency #memorymanagement #rustacean

10

3 Comments

How about categorizing blockchains/ledgers along the following three dimensions? 1️⃣ Integrity & Execution defines who or what is responsible for executing the system and ensuring its integrity 📌 (M) "multi-entity consensus (permissionless)" means a group of independent entities is responsible for processing transactions and ensuring the integrity of the system. The group is permissionless and open in the sense that potentially anyone investing required resources can participate in the consensus algorithm without needing permission (no gatekeeper). 📌 (P) "multi-entity consensus (permissioned)" means a group of independent entities is responsible for processing transactions and ensuring the integrity of the network. The group is permissioned and closed in the sense that the list of validators is managed by one or more gatekeepers. 📌 (S) "single entity" means a single entity is responsible for validating transactions and ensuring the integrity of the system. 2️⃣ Access Control (connect, read-call, write-call) means how permissions for connecting and calling read- or write-functions of the Digital System of Record are managed. Note: Application layer access control is not considered here and can exist on-top of the Digital System of Record, e.g. whitelist managed in a smart contract deployed on an Ethereum Blockchain. 📌 (O) "public (open)" means anyone can connect, read, and write without permission, given transaction fees can be paid (if any). 📌 (M) "mixed (open for read-calls, closed group for write-calls)" means anyone can connect and read, however write-calls are restricted to a closed group managed by one or more gatekeepers. 📌 (C) "private (closed / permissioned)" means the access in general is restricted to a closed group managed by one or more gatekeepers. 3️⃣Programmability 📌 (F) "full" means users of the system can write any Turing-complete programs and execute on the system, given sufficient resources. Turing-complete means that any calculation that can be described algorithmically can be performed. 📌 (L) "limited" means users can write not Turing-complete programs and execute on the system, given sufficient resources. 📌 (N) "not programmable" means users have no possibility to write and execute own programs on the system. ➡️ Categorizing ledgers with a 3-letter code: MOF : Ethereum MOL : Bitcoin POL : Stellar PCF : SWIAT SCN : Target2 … How this categorization explains different public/private and permissioned/permissionless ledger combinations? public/permissionless ≙ MO* public/permissioned ≙ PO*, PM*, SO*, SM* private/permissioned ≙ PC*, SC* private/permissionless ≙ MM*, MC* (a useless paradoxical combination?)

56

6 Comments