Jason Grusauskas (@grussy) header
Jason Grusauskas (@grussy) profile picture

Financial Markets Strategy, Transformation, & AI | Former Goldman Sachs & Bank of America

NYC
Joined October 2015
Photo by Jason Grusauskas (@grussy)
Two announcements from DBS this week caught my attention. First, a partnership with JPMorgan enabling 24/7 cross-border transfers of tokenized deposits. Then, a partnership with Circle providing institutional clients direct access to USDC issuance and redemption. What’s interesting is that DBS is supporting both tokenized bank deposits and stablecoins, rather than betting on one model. As these different forms of digital money develop, the ability to move liquidity between them may become just as important as the instruments themselves. DBS / JPMorgan: dbs.com/newsroom/DBS_and_Kin… DBS / Circle: dbs.com/newsroom/DBS_and_Cir…
13
A landmark day for Anthropic and the AI industry. I'm curious to see what public-market scrutiny reveals about the economics of frontier AI. Enterprise adoption, revenue growth, inference costs and the capital required to support increasingly capable models will all be subject to a different level of scrutiny.
1
1
49
Jason Grusauskas retweeted
show this video to anybody that is an old school programmer that hates AI this is one of the best videos I’ve seen all month.
1,123
6,461
1,337
47,074
4,228,601
The World Bank is exploring stablecoin-based disbursements with Citi and Circle to improve the cost-efficiency, speed and transparency of payments and financial settlements for its clients. What's particularly notable is the stated objective of working alongside or improving existing disbursement arrangements built on correspondent banking infrastructure, rather than replacing them. This is particularly relevant for emerging markets, where moving dollars across borders is often only part of the challenge. Local currency liquidity, conversion and access to domestic payment systems remain just as important. worldbank.org/en/news/press-…
9
Interesting to see the Fed's Michael Barr raising similar concerns yesterday about AI-driven trading, correlated behaviour and broader market participation. federalreserve.gov/newsevent…
What happens when independently controlled AI trading agents all respond to market stress in the same way? In a recent Bank of England speech, Jon Hall raises concerns about AI agents potentially increasing correlated trading behaviour and amplifying market stress. For established electronic trading firms, this is nothing new. Correlated behaviour among market makers, particularly during periods of market stress, has long been a feature of electronic markets. What may be different is the breadth of participation. As AI lowers the barriers to developing sophisticated trading strategies, a much wider range of market participants could deploy agents, potentially relying on similar underlying models. Whether that increases market diversity or creates new forms of correlated behaviour remains to be seen. bankofengland.co.uk/speech/2…
5
What happens when independently controlled AI trading agents all respond to market stress in the same way? In a recent Bank of England speech, Jon Hall raises concerns about AI agents potentially increasing correlated trading behaviour and amplifying market stress. For established electronic trading firms, this is nothing new. Correlated behaviour among market makers, particularly during periods of market stress, has long been a feature of electronic markets. What may be different is the breadth of participation. As AI lowers the barriers to developing sophisticated trading strategies, a much wider range of market participants could deploy agents, potentially relying on similar underlying models. Whether that increases market diversity or creates new forms of correlated behaviour remains to be seen. bankofengland.co.uk/speech/2…
1
1
2
25
A new asset class, a familiar derivatives market problem. CME's proposed GPU-compute futures have been delayed again. There's a clear case for hedging compute costs, but establishing a reliable benchmark in a fragmented, privately negotiated market is another matter. ft.com/content/269b0245-44c4…
17
Jason Grusauskas retweeted
HUGE REVOLUT NEWS 🚨 @NStoronsky confirms he wants Revolut to IPO in 2028. Currently gaining around 100k users a day. I cannot wait for this Cheeky Pint episode. 🍻
6
3
1
92
9,083
Interesting BIS research on the relationship between offshore NDF trading and onshore FX markets. Using transaction-level clearing data across ten emerging-market currencies, the researchers quantify how offshore order flow affects domestic spot rates, and how dealer balance-sheet constraints and restrictions on cross-market trading influence that transmission. Particularly interesting are the findings around India's changes to offshore market access and their effect on the onshore/offshore basis. bis.org/publications/working…
14
Jason Grusauskas retweeted
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5: hackertube.org/ - Excellent recs without any logins (state is kept in local browser session) - Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads - Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version - HackerTube Longs ✊ A bit more on the process: I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing. Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
2,415
4,870
1,831
52,692
5,973,392
Jason Grusauskas retweeted
Today is my last day as White House Chief of Staff.   Serving President Trump in the White House has been the honor of a lifetime. I’m so proud to have been a part of what we’ve accomplished together from securing our border and cutting taxes to strengthening our economy and putting America First. More than anything, I’m grateful for the extraordinary people I’ve had the privilege of working alongside. While my role may be changing, my commitment to the President and the mission never will.
14,530
5,265
1,441
53,165
4,690,223
Beyond the headline adoption figures, one finding from Capgemini's latest financial services research stood out: roughly half of firms cannot consistently demonstrate ROI from their agentic AI initiatives. Part of the challenge lies in how institutions make investment decisions in the first place. AI investments carry considerable uncertainty. Returns are difficult to predict upfront and often require experimentation, iteration and sometimes a change in direction. This calls for a different approach to investment governance, borrowing from venture capital models. Fund initiatives in stages, test assumptions, measure outcomes and release further capital as evidence builds. Some initiatives will scale, others will pivot or stop. That's part of the process. Demonstrating ROI remains essential, but the investment framework needs to accommodate uncertainty rather than demand certainty upfront. capgemini.com/us-en/news/pre…
13
The OCC announced a new supervisory framework for AI in banking today. Rather than treating AI as a separate category of risk, the OCC takes a pragmatic approach, focusing on the activities being performed, their consequences and the degree of autonomy involved. This is particularly relevant as institutions develop agents capable of taking action within business workflows. Interestingly, the framework recognizes that human oversight doesn't necessarily mean a person approving every action. At scale, that can become ineffective. For institutions, this points toward a more practical approach to adoption: building the technical controls and operating models that allow autonomy to be delegated safely, with oversight proportionate to the risks involved. The OCC isn't introducing new requirements here, but it is providing a useful indication of how supervisors intend to approach increasingly autonomous AI systems. occ.gov/news-issuances/news-…
21
Restricting AI coding tools won't stop business users from building software. It just buys firms time to catch up. Goldman's decision highlights a growing challenge: demand is outpacing the infrastructure and controls needed to support it safely. Business-built applications are inevitable. Firms need to figure out how to enable them, not just contain them.
Goldman Sachs has irked some of its staff after it banned most of them from using advanced artificial intelligence tools designed for coding after reviewing productivity and risk concerns. Read more here: bloom.bg/4znpAU5 📷: Michael Nagle/Bloomberg
1
55
Jason Grusauskas retweeted
I’m really curious about this. How much extra juice is there, performance wise, for model-specific silicon. It can’t be that easy to design your own chips, for your own models, but everyone send you be doing it. What’s a good thing to read on all this?
Super excited to announce seven new world-class MAI models today. They represent what we consider a new era in AI designed to keep you in control and on the frontier. First is our text foundation model, MAI-Thinking-1, exceptionally strong on reasoning and SWE tasks. - It’s a 35B active parameter MoE with a 256K context window. Independent human raters on Surge prefer it for overall quality in blind side-by-sides versus Sonnet 4.6, and it’s achieved 97% on AIME 2025, the key measure of its general-purpose reasoning abilities. - It's at 53% on SWE Bench Pro, placing it right alongside Opus 4.6 on one of the toughest coding benchmarks. - And since we co-designed our models with our own silicon, MAI-Thinking-1 is optimized on our MAIA 200 chip. Benchmarking head-to-head against the GB200, we see 30% better performance per dollar as well as a 1.4x performance-per-watt gain when running our MAI models on the MAIA 200 end-to-end. Next is MAI-Image-2.5 and its Flash variant. Two super strong models now at #2 on the leaderboards, surpassing the score of Nano Banana 2 on image editing. Last for now is MAI-Code-1-Flash, our new inference efficient coding model, especially tuned for VS Code and GitHub Copilot CLI. - Code-1-Flash achieves 51% on SWE Bench Pro, despite having just 5B parameters, putting it closer to Haiku in size but cheaper in cost. All of this is the foundation for Microsoft Frontier Tuning. It lets you customize our models to create custom, company-specific agents that only you control. You can make our model, your model. Your data. Your agents. Your moat. Early adopters are already seeing a difference. When we tuned our models for McKinsey’s tasks, MAI delivered the highest win rate, outperforming GPT-5.5 on quality, while being 10x lower on cost. Also really excited to be collaborating with the amazing team at Mayo Clinic to jointly train a new frontier AI model for healthcare. Our announcements today mark another milestone on the road to humanist superintelligence. You can learn more and about our other new models in our latest blog: microsoft.ai/news/building-a…
15
4
58
32,344
Jason Grusauskas retweeted
YES! This paper basically confirms what many of us already suspected: If you want better LLM results without paying for longer outputs or fine-tuning, there’s a concrete, low-effort tip: Duplicate your prompt! Researchers found that repeating the exact same input can dramatically improve performance (up to a 76% gain on specific tasks). LLMs process text left to right, each token can only look at the previous context, never ahead. So when you write a long prompt with context first and the question at the end, the model can rely on that context to answer, but the context was processed before the model even knew the question. This asymmetry is a basic structural property of how LLMs work. Repeating the prompt helps counter this limitation by giving the model a second pass over the full context. There are no new losses to compute and no fancy prompt engineering involved. It’s just a simple structural hack that works across almost every major model they tested. Paper in 🧵↓
41
68
12
412
47,009