
Technology Consultant @ Dart . Prev @moderntreasury @beam_cash @lendbuzz, @onswipe, @ColumbiaEntrep
Boston
Joined May 2013
- Tweets3.6K
- Following1.6K
- Followers539
- Likes7K
Frank retweeted
Never send a biologist to do the work of a computer scientist.
Dawkins doesn't understand that evolution built human computational abilities breadth-first — memory, language, object model, generalization and classification, agency, and so on, all in a primitive state, and then refined them.
Computer science isn't doing that.
It is building human capabilities depth-first. So we have something that emulates human language capabilities to an advanced degree...
... but nothing else.
That's why there is a curious sense of something missing when you talk to Claude or Grok or ChatGPT. It's not some minor errors with use of language itself. Its language capabilities are quite advanced.
What you are detecting instead is the complete absence of these other neural systems, which are what lies behind the use of language in people.
Something that is very glib with language but has no object model, no mirroring ability, no understanding of the ground truth of the universe its in might be able to become president of the United States, or win a Nobel peace prize, but it isn't actually a person.
It's more like a small slice of a person's brain, containing Wernicke's and Broca's areas, and very little else.
We're not used to thinking of people as a collection of systems, but we're going to have to start, because we no longer have the luxury of dividing the universe in human and not, and automatically assuming every human is a person, and every non-human isn't.
You can't evaluate a software neural net as if it were a proto-human, and try to decide on that basis whether it is a person that's allowed to do what we allow people to do.
If you allowed a small slice of brain, containing Wernicke's and Broca's areas, to do things like vote or run for office, then it would be able to appear to do so, but have no actual understanding of what was going on, no coherent model of the universe or the task before it.
This would lead to disaster for any number of issues, such as race relations or the medical industry.
Let me be 100% clear... LLMs are not people.
They are not people now.
They will never be people.
And anyone who thinks LLMs are people is probably not a person, either.
We may someday make something that is a person.
But it will have an LLM, not be one.
Frank retweeted
Desde que vivo acá en El Salvador hay algo que me llama demasiado la atención.
Y es que muchas cosas que para los salvadoreños parecen normales… para un extranjero no lo son en absoluto.
La forma en la que la gente saluda.
La cercanía entre familia.
La facilidad con la que alguien te ayuda, incluso sin conocerte.
Eso en muchos lugares del mundo ya no existe así.
En Europa, por ejemplo, muchas veces todo es más frío, más distante, más individualista.
Aquí siento algo diferente: comunidad.
A veces tengo la sensación de que muchas personas aquí no son conscientes de lo especial que es eso.
Porque cuando uno crece con algo, lo ve como “normal”.
Pero desde afuera… se ve como algo muy valioso.
El Salvador no solo está cambiando en temas de seguridad.
Hay algo más profundo que ya estaba aquí desde antes: su gente.
Su forma de vivir.
Su forma de tratar a los demás.
Su forma de confiar en Dios y en la vida.
Y eso, para mí, vale más que muchas cosas materiales.
Ojalá más personas aquí pudieran ver lo que yo veo.
Porque tienen algo que en muchas partes del mundo ya se ha perdido.
Y eso no tiene precio.
Great read, have a peak into my world
Incredible work by @mikulaja
@kontigo_app was at best a naive puppet for sanctions violations
At worst a knowing participant.
The whole stablecoin industry is on notice
Get your AML and Sanctions controls buttoned up or there will be blowback
It's when not if.
fintechbusinessweekly.substa…
2
211
And they should be understood like normal objects, is what I tell my engineers
Computer programs (even very complicated ones) are physical objects that are made of solid-state chemistry and electrical flows. They are not platonic forms, they are not abstractions, they are normal objects like everything else.
1
435
In case you've been wondering what I've been up to, this is some of July's progress
We’ve been hard at work this past month shipping improvements across the stack. From new APIs to stronger infra, here’s what we rolled out in July:
1. U.S. Virtual Accounts are live
Named accounts are now in production to give individuals and businesses around the world the ability to receive, hold, and send USD. Platforms can now offer branded, banking-like experiences without the overhead of becoming a bank.
2. First - and third-party SWIFT support
Beam now supports both first-party and third-party international wires for USD transfers. Teams can now enable global USD transfers with full flexibility in sender/receiver configurations.
3. ACH Pull is now live
Platforms can programmatically debit funds from users’ U.S. bank accounts. Cash flow becomes easier to control and automate across payins. Instant ACH Pull is also available.
4. Stablecoin-to-stablecoin trades now in Hub
Stablecoin-to-stablecoin trades are now available directly in Hub, our no-code dashboard. Teams now have greater control over moving between stablecoins on major blockchains.
5. More API power for U.S. Virtual Accounts
We added endpoints to create virtual accounts, retrieve transaction history, and access transaction metadata. Builders can now deliver deeper financial UX with less backend work.
Looking ahead, we’re focused on expanding Plaid integrations, making Virtual Accounts more programmable, and simplifying operations for teams scaling with Beam.
3
274
I aspire to this level of conviction and focus
Replying to @anonymemphis
i'm a guy who lives alone with a cat, so arguing about clothes online is nothing for me. this is in my blood; it's my purpose. like asking if babe ruth plays baseball all day.
2
317
Been thinking a lot about FinTech fraud lately and this is super interesting
TL;DR: We built a transformer-based payments foundation model. It works.
For years, Stripe has been using machine learning models trained on discrete features (BIN, zip, payment method, etc.) to improve our products for users. And these feature-by-feature efforts have worked well: +15% conversion, -30% fraud.
But these models have limitations. We have to select (and therefore constrain) the features considered by the model. And each model requires task-specific training: for authorization, for fraud, for disputes, and so on.
Given the learning power of generalized transformer architectures, we wondered whether an LLM-style approach could work here. It wasn’t obvious that it would—payments is like language in some ways (structural patterns similar to syntax and semantics, temporally sequential) and extremely unlike language in others (fewer distinct ‘tokens’, contextual sparsity, fewer organizing principles akin to grammatical rules).
So we built a payments foundation model—a self-supervised network that learns dense, general-purpose vectors for every transaction, much like a language model embeds words. Trained on tens of billions of transactions, it distills each charge’s key signals into a single, versatile embedding.
You can think of the result as a vast distribution of payments in a high-dimensional vector space. The location of each embedding captures rich data, including how different elements relate to each other. Payments that share similarities naturally cluster together: transactions from the same card issuer are positioned closer together, those from the same bank even closer, and those sharing the same email address are nearly identical.
These rich embeddings make it significantly easier to spot nuanced, adversarial patterns of transactions; and to build more accurate classifiers based on both the features of an individual payment and its relationship to other payments in the sequence.
Take card-testing. Over the past couple of years traditional ML approaches (engineering new features, labeling emerging attack patterns, rapidly retraining our models) have reduced card testing for users on Stripe by 80%. But the most sophisticated card testers hide novel attack patterns in the volumes of the largest companies, so they’re hard to spot with these methods.
We built a classifier that ingests sequences of embeddings from the foundation model, and predicts if the traffic slice is under an attack. It leverages transformer architecture to detect subtle patterns across transaction sequences. And it does this all in real time so we can block attacks before they hit businesses.
This approach improved our detection rate for card-testing attacks on large users from 59% to 97% overnight.
This has an instant impact for our large users. But the real power of the foundation model is that these same embeddings can be applied across other tasks, like disputes or authorizations.
Perhaps even more fundamentally, it suggests that payments have semantic meaning. Just like words in a sentence, transactions possess complex sequential dependencies and latent feature interactions that simply can’t be captured by manual feature engineering.
Turns out attention was all payments needed!
3
242
Gold idea and experience
You can start your own company and work for it under an H-1B.
Read that again.
Here's the playbook, and the steps to take in a 60-day grace period
You need to:
- register a US company
- prove company can pay salary
- find someone else to work with (cofounder, board member, advisor, etc)
- have them file your visa petition for you
A bad application would look like:
- company formed last week
- no income noted on "gross income" and "net income" sections
- no product, customers, revenue
So what do you do?
🖥️Throw a website up to describe what the company does. Are you building product? Consulting? An e-commerce store?
Everything is easier if the gov't can understand what your company does and how you fit into it.
👫Find a friend that wants to build with you. They don't need to be:
- American
- on payroll
- shareholders
They will sign forms to represent the company and offer you your H-1B role.
💰The hardest part here is proving a company can pay your H-1B salary. If you can quickly raise some money, or move some of your own money into your company, you can draw some of that out as salary.
I've seen applications with zero revenue and zero money in the bank get approved. It's risky, but it can work.
Don't listen to anyone that makes up a magic number and says something like "you need 3 years of salary in the bank," "you need $150k or more", etc.
💼Keep your job title/ field/ SOC code as similar as possible. Going from a SW engineer to a sales role/ BD guy will be challenging. Maybe doable, but challenging. If you can keep things similar, the application will be more likely to succeed.
Find a legal team that's handled applications like this before. 👋
We talk to people every week that fuck up their cap table, take bad investments, hire someone they didn't want... all because they thought it was necessary for them to get a visa.
Don't do anything that doesn't make sense for you or your new business without taking to an attorney and understanding the risks/ tradeoffs.
**If you're in your grace period/ need to file ASAP:
Rule of thumb: We need to file the application by day 60 from your last paycheck.
Day 1 - register the new business.
- Delaware C Corp, Wyoming LLC, California S-Corp... pretty much any type of business will do, so long as it's registered in the US. Talk to a tax/corporate lawyer about the specifics, but you can always change later to make the taxes hurt less. Tax mitigation is not the priority right now, nor is he merging with some other entity in another country.
This will take somewhere between 3 days and 5 weeks depending on where you are.
T-45 days: Put your website up and figure out who will work with you and sign paperwork for your company. Do not build, sell, hire, yet - you don't have work authorization.
T-30 days: Make sure you have an EIN number assigned to you.
T-25 days: Figure out your job title, salary, worksite location, etc. (Legal team can give guidelines/ ranges for you to work with.)
T-18 days: Get your EIN number approved by the Dept of Labor. (Dumb step, attorney will handle)
T-15 days: File the LCA for your H-1B (attorney will handle)
It should take a week-ish to come back as approved, you can now file the H-1B.
T-7 days: file the H-1B. Technically it just needs to be sent before your 60th day, but don't play with fire. Get it in their hands on day 60 or earlier.
Wait for your approval to officially start work for your own company.
Post-approval: Get yourself on payroll and get to work!
🎊 🍾
363
Hate to be a hater, but he's right. Founder needs to hedge the very real risk of failure on their future finances
1
343
Incredible thread about hitting the next level of things. As a former sprinter and tech exec, I can relate. Folks at the next level always looked like they were operating on an entirely different set of parameters, a different world
Replying to @visakanv
often the reason for this is that the people before the threshold are operating with a set of beliefs/assumptions that are incompatible with life beyond the threshold. My instinct is to say “this maybe doesnt apply to physical things like sprinting” but then, Daniel Chambliss:
1
7
424
Just when I think I've mastered some Claude feature they release another one
It's only been a day since Anthropic dropped Model Context Protocol (MCP) with Claude, like universal plug for AI.
Now AI connects to tools and data insanely easier.
And people can't stop getting creative with it to do your work.
10 wild examples:
2
123
Essays could be written about this pattern
A balanced life has work, friends, family, health, hobbies, etc. An imbalanced life goes very deep in one at the expense of the rest.
I think you want life to have chapters of balance and imbalance. Happiness comes from balance, progress comes from imbalance.
1
112
Very insightful thread. Reminds me of teaching junior developers to see around corners, most of good programming is avoiding writing bad programming
i built an app with ZERO coding experience using AI.
it was the happiest day of my life.
here's why i don't recommend anyone doing the same.
my learnings and reflections:
🧵
1
92
I would highly recommend working for this man if you can swing NYC. He is smart, driven, and fair - the current opp in real estate is also setting up nicely
Hiring full-time analyst/associate, DM:
Location: NYC
Base: $125k
Current support role, made $350k all-in with promote, fees, co-invest last 12-months. If you work, upside there.
$55M closed last 24-months, pure acquisition mode. We aren’t buried managing bad deals bought in 2021.
1
265
There’s always an asset that outperforms, somehow I didn’t think commercial real estate would be the one
Cap rates in commercial real estate have risen somewhere between 150-250bps over the last 24 months.
With borrowing costs now beginning to fall, we are entering a phenomenal buying “moment” akin to 2011-2013.
Expect transaction velocity to increase substantially in 2025.
1
261





















