Benoît Brookens III (Dr. Big Data)

⚡ 8.3 Million Orders / Second — Under 100 Nanoseconds Latency. This is what “production-grade” really means in high-frequency trading systems. The architecture below is a minimal yet realistic C++ HFT Exchange Simulation: Lock-free queues across order flow Multicast UDP market data Trade engine executing microsecond-level algorithms Benchmarks: 🧠 Avg latency — 89 ns 📉 P99 latency — 145 ns ⚙️ Throughput — 8.3M orders/sec 🕹️ Uptime — 99.99% 🚫 Crashes — Zero The design looks simple — but under the hood, every nanosecond is earned through discipline: No heap allocations No locks Pinned CPU cores Cache-aligned structs It’s not about writing clever code — it’s about writing code that disappears at runtime. 🔗 You can explore or clone the full project here: 👉 https://lnkd.in/dsKpcymR Curious: if you were designing this, where would you try to squeeze more performance? 👇 #HFT #Cplusplus #LowLatency #TradingInfrastructure #QuantEngineering #OpenSource

459

13 Comments

Today is a new era of reasoning models. AI systems will never be the same after today. Cerebras Systems is claiming to reach 3000 tokens/second for the new Open Source Model from OpenAI (GPT OSS 120B). I have tested on OpenRouter and reached 697 tokens/s and using Groq it reached 524 tokens/s ... for a REASONING MODEL! To put it in perspective, I asked the same question "What are the latest models released today? Make a table of all of them categorised by provider" using OpenAI OSS 120B and then compared to Qwen 32B. Qwen 32B took 30 seconds to reason and show the answer (other providers) OpenAI OSS 120B on Cerebras and Groq took 0.7 seconds! AI Agents will be blazing fast, helping iterate 30x faster than before!

269

24 Comments

Scaling RL for Trillion-Scale Thinking Model Scaling RL is hard! But this team might have figured out something. They introduce Ring-1T, a 1T-parameter MoE reasoning model with ~50B params active per token. It’s trained with a long-CoT SFT phase, a verifiable-rewards reasoning RL phase, then a general RLHF phase, and introduces three pieces that make trillion-scale RL actually run: - IcePop to stabilize updates - C3PO++ to keep GPUs busy under a token budget - ASystem to unify high-throughput RL stack On benchmarks, it leads open weights on AIME-25, HMMT-25, ARC-AGI-1, LiveCodeBench, CodeForces, and ArenaHard v2. It reaches silver-medal level on IMO-2025 using only natural-language reasoning.

124

5 Comments

Cross-Attention for Cross-Asset Applications In the previous blog post (https://lnkd.in/g8NRYAXY), we saw how we can apply self-attention transformers to a matrix of time series features of a single stock. What if we have a portfolio of many stocks whose returns we want to predict, or whose allocations we want to optimize? That’s where cross-attention transformers come in. The purpose of cross-attention is that a feature in an asset i may be relevant as a context to a feature of an asset j.  Read more at https://lnkd.in/g7PkjbkQ ...

163

4 Comments

DeepSeek, a Chinese AI startup founded in 2023 as a side project of a Hedge Fund, has made a bold entrance into the global AI arena with the unveiling of its R1 model on January 20, 2025. Claimed to rival industry leaders like OpenAI’s ChatGPT and Google’s Gemini, R1 has drawn attention for its efficiency: trained on only 2,000 Nvidia chips and a reported $5-6 million budget, far below the infrastructure investments typical of Western AI leaders. By fully open-sourcing R1, including its model weights, DeepSeek has ignited debates around democratizing AI, the transparency of development costs, and how innovation can be achieved with leaner resources. The announcement has already unsettled financial markets, with major tech stocks like NVIDIA seeing significant declines, prompting reflection on how the U.S. maintains its leadership in this critical sector. R1’s debut is a moment of both challenge and opportunity for the U.S. On one hand, its rapid adoption (currently the top download in the Apple App Store) has disrupted assumptions about resource-intensive AI development. On the other, concerns about cybersecurity, transparency, and potential state support highlight the complexities of global competition in this space. For American innovators, there are critical lessons to learn: the importance of balancing high-budget investments with breakthroughs in efficiency, fostering robust open-source ecosystems, and ensuring policies and infrastructure that enable sustained U.S. leadership. As global competitors challenge the status quo, the story of DeepSeek underscores the critical need for the U.S. to remain vigilant in safeguarding its leadership in AI innovation; by relentlessly adapting, investing strategically, and fostering an environment that ensures its position as the driving force shaping the next era of technological progress in AI and beyond. https://lnkd.in/e9PGTXTV

34

2 Comments

How is BPE different from Byte Level BPE? I'm Learning about Tokenization, and this week I read, implemented two Research Papers. -> "Neural Machine Translation with Byte-Level Subwords" by Changhan Wang , KyungHyun Cho , Jiatao Gu -> "Neural Machine Translation of rare words with subword units. " by Rico Sennrich, Barry Haddow , Alexandra Birch BPE Introduced in "Neural Machine Translation with Byte-Level Subwords" which worked in character level of sub words and Byte Level BPE which is enhanced version of BPE Introduced in "Neural Machine Translation of rare words with Subword units", the goal was to solve the issue of Rare subwords (Sometime some words was not used in training time but comes in inference ) than it face issue. BPE worked on Character level, but where Byte Level BPE works in Byte-based subwords that convert the words into bytes ( 1 to 4 ) and then used in BPE, which helps to understand the new words at inference time, and also it performs well in Multilingual. --- BPE - Byte Pair Encoding --- -> Character level on sub words. -> Find most common words and then merge it. -> Face problem with new words and Non-english languages. --- Byte Level Byte Pair Encoding --- -> Byte level ( 256 Bytes ) on sub-words. -> Convert sub-words into Bytes, then Bytes go to BPE. -> Now, Same as BPE. -> Perform well in Rare Sub-words, Non-English Languages. In a few days, I will publish it with a Case Study.

19

2 Comments

#tbt We walk through a high-frequency liquidity-taking strategy and show how to simulate and forward test it using just 114 lines of Python. This strategy is designed to scale to $100M+ notional per day - and now comes with full sample code. Good for educational purposes, quants looking to stress-test execution logic or build their first systematic trading loop. #HFT #QuantFinance #AlgoTrading #Databento #MarketData

167

12 Comments

𝐕𝐢𝐬𝐮𝐚𝐥 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐓𝐨𝐩𝐢𝐜 #121: Induction Heads - How Transformers Learn Patterns. We often talk about attention heads, activations, and token interactions in large language models, but these remain abstract for most people. Sharing an interactive tool - Induction Spotlight, this tool visually explains how transformer models (like GPT) perform induction, a core capability behind their ability to complete patterns and generate coherent text. You’ll see attention heads in action, highlighting how they lock onto repeating sequences, effectively learning the "rules" of continuation from just a few examples. This tool is built on the Mistral-7B model with real token-by-token attention flows and provides intuitive visualizations of model internals. Link is in the first comment

82

1 Comment

𝐀𝐫𝐞 𝐋𝐋𝐌𝐬 𝐎𝐯𝐞𝐫𝐤𝐢𝐥𝐥? NVIDIA’s latest research makes a simple — but important — point: 𝑴𝒐𝒔𝒕 𝑨𝑰 𝒂𝒈𝒆𝒏𝒕𝒔 𝒅𝒐𝒏’𝒕 𝒏𝒆𝒆𝒅 𝒍𝒂𝒓𝒈𝒆 𝒍𝒂𝒏𝒈𝒖𝒂𝒈𝒆 𝒎𝒐𝒅𝒆𝒍𝒔. 𝑻𝒉𝒆𝒚 𝒏𝒆𝒆𝒅 𝒔𝒎𝒂𝒍𝒍, 𝒔𝒑𝒆𝒄𝒊𝒂𝒍𝒊𝒛𝒆𝒅 𝒎𝒐𝒅𝒆𝒍𝒔. They need small, specialized models. Why? • 🔧 Agentic tasks are often narrow and repetitive (API calls, code gen, orchestration). • ⚡️ SLMs offer 10–30x cost savings on inference. • 🔒 Easier to fine-tune and control, with less risk of hallucinations. • 🧩 Naturally modular — multiple SLMs can serve distinct functions. In real-world benchmarks: 𝟒𝟎–𝟕𝟎% 𝐨𝐟 𝐋𝐋𝐌 𝐜𝐚𝐥𝐥𝐬 𝐢𝐧 𝐜𝐨𝐦𝐦𝐨𝐧 𝐚𝐠𝐞𝐧𝐭 𝐟𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤𝐬 𝐜𝐨𝐮𝐥𝐝 𝐬𝐡𝐢𝐟𝐭 𝐭𝐨 𝐒𝐋𝐌𝐬. At current cost curves, continuing to run LLMs for agentic workloads looks less like engineering — and more like avoidable overhead. 👉 Where do you stand? #AI #AgenticAI #SLMs #LLMs #NVIDIA #AIInfrastructure #FutureOfAI #CostEfficiency

37

9 Comments

🦄 Hot on the heels of Quantum Systems, TEKEVER have become Europe's latest UAV unicorn, after their most recent funding round took they valuation about £1b. The funding round was led by existing investors, including Ventura Capital, Baillie Gifford, and the NATO Innovation Fund. This comes as the business also announced plans for a £400m investment in the UK over the next five years through their OVERMATCH program which is set to create 1,000 jobs across the UK. The program aims to enhance the UK's defence industry by establishing Centres of Excellence for Autonomy, expanding testing infrastructure, scaling production of UAS platforms like the AR3 and AR5, and fostering partnerships across the defence ecosystem. Hats off to Ricardo Mendes and the team! Full press release: https://shorturl.at/SlgZe

49