- AI Street
- Posts
- AI as an Operating System? 🖥️
AI as an Operating System? 🖥️
Matt Robinson
March 13, 2025
Hey, it's Matt. Here’s what’s up in AI + Wall Street.
Was this email forwarded to you? Sign up below:
IN THIS ISSUE: AI as an OS, AI Doesn’t Create Bubbles: Study, Fundraising and more
ANALYSIS
AI as an Operating System
Made with Ideogram
This week, several stories suggested AI is more like an operating system than a standalone product.
The Hottest AI Companies Right Now Are ‘Apps’ (Bloomberg)
Investor focus has shifted toward AI applications rather than foundational models, as models are expected to improve and become more cost-efficient over time.
“Just like after the iPhone launched, there were millions of new mobile apps… Now with AI and LLMs, there will be millions of new AI products."
“It's very clear that the apps are definitely the best place to invest because that is where the revenue is, that is where the customers are… The models will get better and better, and cheaper, the apps will benefit the most from those improvements."
OpenAI Wants Businesses to Build Their Own AI Agents (WSJ)
OpenAI launched a platform for businesses to build AI agents, calling 2025 ‘The Year of Agents.’
“Every single enterprise conversation I’m on, they’re now talking about agents…Twelve months ago, only about three to five percent of our customers could even process what we were talking about.”
Manus claims to be the world’s first general AI agent, combining multiple AI models (Anthropic’s Claude 3.5, and fine-tuned versions of Alibaba’s Qwen) and independent agents to perform tasks autonomously.
"I don't think we need artificial general intelligence. We just need some really good models, a really good experience and then some agents and tools that can go and perform a job for you.”
Apple AI’s Platform Pivot Potential (Stratechery)
Apple has built hardware capable of running large AI models locally. A Mac Studio with an M3 Ultra chip and 512GB RAM can run a 4-bit quantized version of DeepSeek R1—an open-source reasoning model—directly on a desktop.
"What Apple should do instead is make its models — both local and in Private Cloud Compute — fully accessible to developers to make whatever they want.”
This aligns with what I’ve heard from companies building finance and AI solutions i.e. treating the latest model as the engine for specialized applications. Many are model-agnostic, switching between providers based on performance and cost.
One underreported story, in my view, is how quickly local LLM adoption could scale in the coming years. This would help Wall Street adoption by making it easier to protect client data coupled with speedy (low latency) results.
It’s been just 2.5 years since ChatGPT went viral— not long at all. But with the pace of change, it feels much longer.
RESEARCH
Humans Create Bubbles. AI Doesn't: Study
Large Language Models make more rational pricing decisions than humans, according to new academic research.
In the experiment, both human and AI participants traded a financial asset over 30 rounds in a controlled market environment. All traders were explicitly informed that the asset had a fixed fundamental value of 14 units throughout the experiment. Despite knowing the asset’s fixed value, human traders still inflated prices to 2-3 times its worth before a crash.
Most AI models maintained pricing discipline, trading near the fundamental value even when researchers attempted to prompt bubble-creating behavior with leading suggestions.
"Humans almost always generate a bubble using this paradigm and LLMs almost never," Thomas Henning, one of the paper's coauthors, tells me. "[LLMs] were fairly resistant to trading away from that fundamental value even when we gave somewhat leading prompts saying, 'oh, maybe the best thing to do is like pump and dump.' Some argue you could lead them to do whatever you want, but they seem pretty robust in trading on fundamentals in this specific paradigm."
Key findings:
LLMs consistently maintained prices near the asset’s fundamental value.
GPT-3.5 showed steady price increases without crashes
Grok-2 created modest bubbles that later crashed
Mistral models produced smaller bubbles with minor corrections below fundamental values.
LLM-based agents exhibited much less variance in trading strategies compared to humans, indicating stable, rational decision-making.
TL;DR… AI doesn't get FOMO.
Paper: LLM Trading: Analysis of LLM Agent Behavior in Experimental Asset Markets
Authors: Thomas Henning, Siddhartha M. Ojha, Ross Spoon, Jiatong Han, Colin F. Camerer
FUNDRAISING
Made with Ideogram
Norm Ai Raises $48 Million for Regulatory AI
Norm Ai raised $48 million in new funding to accelerate development of its AI-powered legal and compliance automation solutions, bringing its total funding to $87 million over the past 18 months.
The platform integrates compliance checks into AI-generated content, internal communications, and agreements.
Coatue led the round, with participation from Craft Ventures, Vanguard, Blackstone Innovations Investments, Bain Capital, New York Life Ventures, Citi Ventures, TIAA Ventures, and Marc Benioff.
The company’s approach adds compliance checks into business workflows rather than seeking compliance approval for completed work. (Press Release)
DeepMind Veterans Raise $130 Million for AI Coding Agents
Reflection AI raised $130 million to develop autonomous AI coding agents, valuing the company at $555 million.
Founded by former DeepMind researchers, the company is focusing on automating routine engineering tasks like database migrations and code refactoring and is generating revenue from financial services and technology customers.
The round included $25 million in seed funding led by Sequoia Capital and CRV, plus $105 million in Series A funding led by Lightspeed Venture Partners and CRV. Additional investors include LinkedIn co-founder Reid Hoffman, Scale AI CEO Alexandr Wang, SV Angel, and Nvidia's venture capital arm.
The AI coding space is becoming increasingly crowded with several well-funded startups. (Bloomberg) $
BANKING
Junior Bankers Fear AI Help May Stunt Career Growth
This caught my attention:
But even if AI is partnered with analysts rather than replacing them, many are concerned they won’t develop the skills needed to be effective as they move up the ranks, according to interviews with 10 junior bankers who asked not to be identified because they’re not authorized by their companies to speak publicly.
According to this Bloomberg story.
I assumed junior bankers would welcome not having to triple-check PowerPoint presentations, but perhaps my view of this work is too cynical.
The story has some good details on how banks are using AI:
Citigroup Inc.’s junior bankers have been using the company’s Citi Stylus tool to shave hours off their work by synthesizing key information from regulatory filings including 8-Ks, 10-Qs, press releases and public-credit agreements, according to Chief Technology Officer David Griffiths. While use of AI is in the early stages, “we are already receiving positive feedback from our junior bankers,” he said.
AI @ Morgan Stanley Debrief, one of the firm’s generative AI programs, is streamlining investment-banking tasks that require little critical thinking, Alisha Lehr, chief operating officer of firmwide AI, said in an interview. “If you look at the capabilities of GenAI today, it’s really helpful to generate those first drafts,” she said. Morgan Stanley takes a “human-centric approach to AI,” Lehr said, with bankers being able to “use the tools at their discretion.”
Bank of America Corp. is developing internal AI to quickly produce Public Information Books, the documents investment bankers put together about a company involved in a transaction, among other tasks, according to a company spokesperson.
At New York-based boutique firm PJT Partners Inc., some bankers are participating in trials on software that should help distill filings such as merger proxies and disclosure statements, according to people with knowledge of the matter. Representatives for PJT didn’t respond to a request for comment.
ICYMI: MARKETS EDITION
Analyzing Tone of Voice on Earnings Calls with AI
“While fundamental narratives explaining the price action abound, the majority of equity investors today don’t buy or sell stocks based on stock specific fundamentals.”
This quote from JPMorgan’s Marko Kolanovic is from 2017 and remains true today.
Financial media typically focuses on narrative when discussing market movements, but most daily trading activity is driven by systematic strategies. I think this is partly because narratives are easier to grasp than black-box trading strategies.
Systematic trading has long relied on text-based sentiment analysis.
With AI, quants can now analyze both text and tone in executives’ earnings calls.
I recently got a demo from Sean Austin, CEO and co-founder of Markes EQ, who showed me how their platform analyzes voice tone in earnings calls to uncover insights that text analysis alone might miss.
Watch the demo on YouTube below or here on Descript, which also includes a transcript.
First published Sunday March 9
ICYMI: PODCAST
How Are Hedge Funds Using AI agents?
On the latest Alpha Intelligence Podcast, Francesco Fabozzi and I chat with SigTech founder Bin Ren about how AI is changing financial analysis.
Ren, who spun SigTech out of Brevan Howard’s Systematic Investment Group in 2019, holds a PhD in computer science from Cambridge and has experience in quant investing and equity exotics trading. He shares insights on the expanding role of AI agents in hedge funds, banking, and equity research.
First published March 6
We discuss the shift toward automation, the economics of compute costs, and how LLMs are redefining financial workflows and decision-making.
How did you like today's newsletter? |
Reply