Matt Robinson

@richardcraib runs one of Wall Street’s most unconventional business models: a crowdsourced hedge fund. He also counts JPMorgan as his biggest backer. Now, he's retooling @numerai for AI agents. This was a fun interview. ai-street.co/p/the-hedge-fu…

@FundamentEdge Got it. I'm still surprised how much the conversation is still about the the latest capabilities and not about evaluation/ figuring out if its outputs are accurate.

@robinsonmatt it's not there one-shot reliably yet. Requires a thoughtful validation process. But emerging MCPs help a lot in my experience (in past, LLMs wanted to web scrape financial data which is a really bad idea)

It's really remarkable how fast AI tools for Excel have evolved. Even three months ago I found them almost completely unusable. Today, I was able to update my Uber model for the last four quarters in a fraction of the time, accurately, even when I consider the time I spent de-bugging and validating the key inputs. The three big unlocks for me were creating my own skills files, which are recipe cards encoding an incredibly detailed dissection of every step of the financial modeling process (put together in an 86 page document then crafted into six distinct modeling skills...unfortunately, I won't be sharing this at this time, but will consider in the future), connecting the Daloopa MCP to Claude in Claude Excel for accurate data, and creating a validation space in Perplexity Computer to do final checks and de-bugging. (I am not sponsored by either Daloopa or Perplexity, or any vendor for that matter) Obviously this AI augmented process is only valuable to the extent that it is 98%+ accurate and 100%+ accurate on critical metrics. Validation has to be a systematic process blending coding tools and human validation checklists (i.e. hand checking key model variables and understanding where in the model there is tolerance for mistakes, and where there isn't). But the ability of new LLMs to read & analyze models (particularly GPT 5.4) and the rise of Agentic Workspaces like Perplexity Computer to route tasks to the right LLMs seems to be resulting in big progress here. Really exciting stuff. I have been a huge skeptic here...Excel-based models are the foundation of institutional decision making, and they are no place for AI slop. With the technology improving, particularly workflows around systematic validation, that skepticism is melting.

@FundamentEdge This is really the core challenge of combining probabilistic LLMs with deterministic software. A lot of current work is about building scaffolding between those layers, but there’s no settled architecture yet. I think it eventually gets there, but it’s hard to say how quickly.

This is one of the biggest sticking points on AI Excel that I'm trying to understand. 73% accuracy is progress, but is it useful for anything at all? We were on a vendor call last month and the vendor bragged of hitting 65% accuracy in Excel and Andrew Carr and I texted "an analyst who is 65% accurate in Excel is 100% fired". Why is AI Excel only 60-70% accurate? Are these issues fundamental or solvable? > Is MCP fundamentally too brittle to get to 99% accuracy? > Is the data layer clean enough to hit 99% accuracy (i.e. there's a reason why hedge fund analysts don't start their models with a Bloomberg download) > Are the foundation models powerful enough to handle the multi-modal (filings, PRs, investor decks, data supplementals), multi-document, "needle in a haystack" issues for LLMs? Context windows have grown, but they are still not large enough to capture all of the documents and files for one ticker (letalone a coverage universe) > Is the commercial opportunity large enough for foundation labs to build RL environments for public equity modeling, as they are doing on investment banking modeling? Does the "march of 9s" on AI Excel take 6 months or 6 years? Driverless cars took 13 years from DARPA Urban challenge to first Waymo. These are legit questions. I don't know. I also don't really trust public evaluation sets (i.e. LLM's win physics competitions...then you learn the LLM trained on the physics competition test bank lol). The real questions in investment research modeling are out of sample questions (i.e. how to model SAAS retention in a Claude-world...there is no prior on which to rely). So I am building my own evaluation set. 100 use cases ranging from simple (input 3 statements from 10-K to AMZN model) to complex (model GE split/spin). Am I wasting my time? 36 months form now, are we still only at 80% accuracy in AI Excel? These are questions, now answers - love your takes in replies or DM!

...quantitative finance. Takeaway: LLMs can be a Swiss Army knife in daily life, but for mathematical analysis, specialized agents may be the sharper tool. ai-street.co/i/1835820

The authors argue this setup improves analytical rigor and helps mitigate behavioral biases like overconfidence. While limited in scope and not a full portfolio optimizer, the study suggests specialized, debating agents may prove more reliable than general models for ...

𝗕𝗹𝗮𝗰𝗸𝗥𝗼𝗰𝗸 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵𝗲𝗿𝘀 𝗗𝗲𝘃𝗲𝗹𝗼𝗽 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁 𝗦𝘆𝘀𝘁𝗲𝗺 𝗳𝗼𝗿 𝗦𝘁𝗼𝗰𝗸 𝗣𝗶𝗰𝗸𝘀 Instead of relying on one frontier model, BlackRock built three AI “agents” that mimic different analyst roles:

I'm delighted to be featured in the AI Street newsletter!

Where chatbots fail, @HudsonLabs delivers. Find out why the Co-Analyst is getting rave reviews from institutional investors. Thank you for the feature @robinsonmatt. Highly recommend subscribing. ai-street.co/p/separating-f…

@ZekeFaux haha gotta have b roll of "the Town" and have bill burr show up for some reason.

@robinsonmatt As long as the thieves are pats fans, I stand ready to tell their story

Oh man I want the tick tock of this cc @ZekeFaux NFL warns of burglary groups targeting players apnews.com/article/nfl-se…

Great to catch up @TedMerz!

"The lines are blurring between quantitative and discretionary investment strategies, with platforms like Bigdata.com bridging the gap."