Karin Kimbrough

Takeaways from NVIDIA's earnings Mgmt. guided Q1 Rev above at the midpoint, Gross margins below, with improvement post-Blackwell ramp to mid-70s later in 2025 despite complex configurations. Analysts were largely upbeat following another strong print and guide, with Blackwell strength a key highlight, dispelling prior supply worries. Although Q1 gross margin outlook, networking weakness, and China headwinds caused slight concerns, it is anticipated that gross margins recover through the year with networking rebounding as Blackwell ramps. Furthermore, mgmts. reassuring commentary on AI reasoning compute needs addressed some apprehension around DeepSeek and ASICs.

1,263

81 Comments

Top AI Models for Stock, ETF, and Crypto Trading Ranked in New Benchmark A group of academic researchers, led by Stevens Institute of Technology in collaboration with Columbia University, Harvard University, and The Fin AI, released a new paper to evaluate different AI models for stock, crypto, and ETF trading. The framework, called InvestorBench, tested 13 different large language models on their ability to trade financial assets. The system tests each AI model's performance across different time periods: stocks from mid-2020 to mid-2021, cryptocurrencies throughout most of 2023, and ETFs from mid-2019 to late 2020. The researchers built what they call a "memory system" based on an earlier framework called FINMEM to help AI make trading decisions, incorporating both memory storage and reflection mechanisms that help explain trading decisions. During an initial training period, the system fed each AI model market data and financial news to spot patterns between information and subsequent price movements. The system has three memory layers—short-term (14 days), medium-term (90 days), and long-term (365 days)—mimicking how human traders might weigh recent versus older market information. During the actual testing phase, the thirteen AI models used this memory system to make daily trading decisions across three types of assets. Each morning, the AI models would analyze current market conditions and news, compare these to patterns stored in their memory, and issue a simple command: buy, sell, or hold. The best-performing AI model for stocks, Alibaba’s Qwen2.5-72B-Instruct, achieved a 46% return with strong risk-adjusted performance compared to a 34% return for a simple buy-and-hold strategy. For cryptocurrency trading, OpenAI’s GPT-o1-preview led the pack, while Writer’s Palmyra-Fin-70B showed the best results in ETF trading. "Large language models are amazing. They are good at many things, but for some tasks, they're not so good so that's why we choose to do an agentic framework, which combines different components together,” Jimin Huang, founder of Fin AI, told me. The group was established in February 2024 to foster transparency and collaboration in an industry where companies often resist sharing their AI performance metrics. Hedge funds, banks, and startups have approached Fin AI to learn from their research, but these conversations often become one-way streets, with firms unwilling to share their own findings or performance data in return. Thanks to Jimin Huang and Haohang Li for walking me through their research! Link to the paper below. Follow me (Matt Robinson) and AI Street for more on LLMs + investing. 

214

8 Comments

The fight for AI dominance continues to move at a relentless pace. Events this week have underscored compute themes I have been writing about recently. Compute infrastructure has become AI's existential bottleneck. It is the new moat. Here's OpenAI President Greg Brockman on CNBC talking about the $100 billion NVIDIA investment: "You want every person to be able to have their own dedicated GPU, right? So you're talking on order of 10 billion GPUs. We're going to need this deal we're talking about, it's for millions of GPUs. We're still three orders of magnitude off of where we need to be. We're heading to this world where the economy is powered by compute, and it's going to be a compute-scarce one." CEO Sam Altman published a short but sweeping manifesto called “Abundant Intelligence": “To be able to deliver what the world needs—for inference compute to run these models, and for training compute to keep making them better and better—we are putting the groundwork in place to be able to significantly expand our ambitions for building out AI infrastructure." Meta CEO Mark Zuckerberg discussed his infrastructure spending plans a few days ago on a podcast: “…if we end up misspending a couple of hundred billion dollars, I think that that is going to be very unfortunate, obviously. But...the risk is higher on the other side. If you build too slowly and then super intelligence is possible in 3 years, but you built it out assuming it would be there in 5 years, then you’re just out of position on what I think is going to be the most important technology that enables the most new products and innovation and value creation in history.” None of these players can afford to be cautious. Waiting means falling irrecoverably behind. However, the risks of this "Damn the torpedoes, full speed ahead" can not be understated. Bain & Company Capital this week released its sixth annual Global Technology Report, which calculates that these companies will need $2 trillion in new annual revenue by 2030 to pay for this compute, but currently appear to be on pace to fall about $800 billion short. At the same time, a new Financial Times analysis of SEC filings and transcripts for S&P 500 companies raised fresh questions about AI’s short-term utility in the enterprise, concluding: “The biggest US-listed companies keep talking about artificial intelligence. But other than the ‘fear of missing out’, few appear to be able to describe how the technology is changing their businesses for the better." So, is this a bubble? This remains the wrong question. By using the lens of discontinuity, we can avoid focusing solely on the short-term bubble narrative, which misses the larger truth: we're witnessing the early stages of a general-purpose technology deployment on par with electricity or the internet. The question isn't whether AI will transform the economy. It's which companies will capture the value. And when.

62

5 Comments

Today at 1:15 PM EST learn more about Synthetic Data for Portfolios: A Throw of the Dice Will Never Abolish Chance Charles-Albert Lehalle, Head of Quantitative R&D at Abu Dhabi Investment Authority (ADIA), will discuss how generative models are popular in image processing, sound processing, texts and molecular biology. Engineers and researchers are tempted to use them to generate synthetic returns to optimise portfolio. In this paper we show it cannot work with the standard approach, and especially that for portfolio construction, generating more data does not give access to lower confident intervals. He will explain the deep reasons for that; they have two origins 1. Generative models have a bias  2. Portfolio construction is particularly exposed to their bias. We also propose a methodology to evaluate Generators of returns for portfolios, so that researchers in quantitative finance will collectively progress like it has been in other fields. To learn more, visit: https://lnkd.in/eQjjaD4Z #arpm #quantbootcamp #syntheticdata #quantitativefinance #portfolioconstruction #generativemodels #financialengineering #quantresearch #machinelearning #investmentresearch #riskmanagement

40

Why do even the smartest language models hallucinate? OpenAI’s latest research shows that #LLMs #hallucinate not because they lack information, but because they’re trained and evaluated in ways that reward confident guessing over admitting uncertainty. A striking example: when asked about the title of co-author Adam Kalai’s PhD dissertation, the model confidently produced three different answers — all wrong. It did the same with his birthday, inventing three dates, none of which were correct. If models can fabricate details about their own author, imagine the risks when they’re applied without the appropriate understanding of how to use AI in #law, #finance or #medicine. The full paper digs into why this happens and how we might change training to fix it: “Why language models hallucinate”: https://lnkd.in/eikcrXJK 👉 When was the last time you caught a model making something up?

38

3 Comments

This week, my article titled "The Open Source Revolution of DeepSeek" was published in the Hong Kong Economic Times (in Chinese). It analyzes how the "Open Source" model has contributed to the success of DeepSeek amid U.S. technological restrictions against China, including GPU export bans, SPE sales restrictions, investment curbs, and even the inaccessibility of ChatGPT. The Open Source model offers full transparency of source codes, allowing the public to modify and share them. This has helped DeepSeek quickly gain trust from users worldwide. Additionally, it attracts developers globally to collaborate and build a new ecosystem together—something that is not easily achievable for a China-based company. Open Source is also a highly cost-efficient way for DeepSeek to continue enhancing its models. As a result, DeepSeek has not only provided an innovative solution to train its large language models (LLMs) at a much lower cost but has also successfully leveraged the Open Source model to overcome U.S. technology restrictions. While tech firms in Silicon Valley debate whether open-source AI will become mainstream, DeepSeek has already democratized AI by making it more affordable. The adoption of AI in software development should accelerate as more AI application layers are built on top of existing frameworks. In fact, I believe that the entire value chain—including homegrown semiconductor companies, foundries, cloud service providers, AI infrastructure developers, robotics manufacturers, and even end users like us—will ultimately benefit from this shift. https://lnkd.in/gEURvGWT #DeepSeek #opensource #techwar #AI #AIapplication #innovation

55

2 Comments

🚀 Parameter-Efficient Fine-Tuning of LLMs just received another boost 🚀 Excited to share the new research by the Quantum-Inspired Algorithms Team, Global Technology Applied Research, JPMorganChase 📄 MetaTT: A Global Tensor-Train Adapter for Parameter-Efficient Fine-Tuning https://lnkd.in/dUinGDKd MetaTT introduces a unique way to fine-tune large pre-trained transformers using a single, shared Tensor-Train (TT) adapter. Unlike widely-used LoRA, which fine-tunes each transformer sub-module - queries, keys, values, projections, and MLPs - independently using low-rank matrices, MetaTT captures shared structural patterns across all submodules by indexing modes like layer, type, heads, and tasks within a single TT core. 🔍 Why MetaTT? • Significantly reduces final adapter parameter count: up to 20× fewer parameters than LoRA while achieving comparable accuracy on standard language modeling tasks. • Scales better with model size - when tested from RoBERTaBase to RoBERTaLarge. • Supports multi-task adaptation across GLUE datasets with a shared TT representation by introducing a separate TT core to capture dependencies across different tasks. • Introduction of DMRG-inspired rank-adaptive training, a technique from quantum many-body physics, to optimize TT ranks during training. 📊 We benchmark MetaTT against full fine-tuning, LoRA, VeRA, and previous Tucker-Tensor based fine-tuning method LoTR (https://lnkd.in/dB9fexHr). Among all methods, MetaTT consistently provides the best parameter-efficiency–accuracy trade-off, with significant gains in model compression and adapter reuse. 👏 Big kudos to the team for pushing the boundaries of Quantum-Inspired techniques in deep learning! Co-authors: Javier Lopez Piqueres Pranav Deshpande Archan Ray Mattia Jacopo Villani Marco Pistoia Niraj Kumar #LLMs #NLP #Transformers #LowRankAdaptation #TensorTrain #FineTuning #QuantumInspired

81

NEW: Deepseek just dropped ANOTHER new open-source model. Janus-pro-7B. It's multimodal (can generate images) and beats DALLE-3, and Stable Diffusion across GenEval and DPG-Bench benchmarks. This follows all the R1 hype. The competition looks cooked. But we've seen this movie play out several times now. Someone else comes out swinging with multiple drops. I'm already seeing start-ups adopting parts of the open-source models for their stack. Especially R1 which is exceptionally good at selecting other foundation models for reasoning tasks. Nvidia is down 3% today on the view that this all somehow threatens chip demand because the models are small and efficient. I think something else. I think this makes models small enough to live in your mobile phone and be high quality. This isn't a China vs USA thing. It's a we're-living-through-history thing.

107

18 Comments

The current transparency measures of the EU code of practice for GPAI models don't provide enough information for downstream players to comply with the AI Act's high-risk requirements, according to experts of the appliedAI Initiative GmbH. While debates around copyright and societal risks have dominated the conversation, this critical area has received far less attention. The researchers caution that unless this gap is addressed in the final version, we risk either weakening compliance standards or slowing down technological adoption. Dr. Till Klein.

104

9 Comments

An entire trading ecosystem built around Nvidia alone is poised for a test again. Not just options, there is a huge ETF ecosystem around $NVDA that are increasingly used by traders to make leveraged and inverse bets. Here's a rundown.

22