Richard Pham

shiii 😂

Our engineers just be building claude code alternatives as side projects when they feel like it

@aduermael Crazy side project

I've been working on this humble Claude Code alternative. In a nutshell: containerized by default, multi-provider (Anthropic, OpenAI, Gemini & Grok so far), self-building dev environments & 100% open-source, 100% Go. The repo is brand new, only 1 ⭐️, 🥲.

We don’t need to fix this Accomplishing something big should take tremendous sacrifice Most sacrifice their families and marriages first, which to me is a far greater sin Ofc success requires some balance point, but you should be prepared to dip hard into the red

best way to keep up with new AI startups: drive to SFO and read the billboards on 101

@Fer_Populi @bilawalsidhu i suggest you diversify your news sources a bit :). reading is a useful skill! reuters.com/world/middle-e…

@phamrich_ @bilawalsidhu it was iranian missile, haven't you see the videos at x a week ago?

Probably the most current look at Palantir’s maven smart system software. Here’s the DoW’s Chief AI officer showing how it works:

ok are we sleeping on Fremont? 😂

@nicochristie soft launch

Some good nuggets here Spreadsheets are the computation layer, the reasoning layer, and the presentation layer in way no other app combines The first mass use case for computers It's been 40 years since the last big shakeup brb for Shortcut v1 ❤️

@andrewchen humans do not work headlessly. We need to see the state on a grid-like surface so we can inspect local relationships. that is, until we start trusting AI blindly

@phamrich_ this is what an IDE and DB browser and whole slew of data science tools PLUS whatever LLMs can now generate on the fly can use

prediction re the end of spreadsheets AI code gen means that anything that is currently modeled as a spreadsheet is better modeled in code. You get all the advantages of software - libraries, open source, AI, all the complexity and expressiveness. think about what spreadsheets actually are: they're business logic that's trapped in a grid. Pricing models, financial forecasts, inventory trackers, marketing attribution - these are all fundamentally *programs* that we've been writing in the worst possible IDE. No version control, no testing, no modularity. Just a fragile web of cell references that breaks when someone inserts a row. The only reason spreadsheets won is that the barrier to writing real software was too high. A finance analyst could learn =VLOOKUP in an afternoon but couldn't learn Python in a month. AI code gen flips that equation completely. Now the same analyst describes what they want in plain English, and gets a real application - with a database, a UI, error handling, the works. The marginal effort to go from "spreadsheet" to "software" just collapsed to near zero. this is a massive unlock. There are ~1 billion spreadsheet users worldwide. Most of them are building janky software without realizing it. When even 10% of those use cases migrate to actual code, you get an explosion of new micro-applications that look nothing like traditional software. Internal tools that used to live in a shared Google Sheet now become real products. The "shadow IT" spreadsheet that runs half the company's operations finally gets proper infrastructure. The interesting second-order effect: the spreadsheet was the great equalizer that let non-technical people build things. AI code gen is the *next* great equalizer, but the ceiling is 100x higher. We're about to see what happens when a billion knowledge workers can build real software.

@Clearingfog_ Damn that's crazy...anyway have you tried @tryshortcutai ?

Too many VPs and Associates have trained this thing:...already pushing back.

there's this false dichotomy I see a lot: either my agent does a task or I do it. real-life tasks are highly complex and require judgement from a knowledge worker's taste and expertise (for now). if you're using AI, you should really be asking: where do I need to insert a human in the loop (me) to ensure there's no deviation from my optimal path/output

@FundamentEdge I understand what you're trying to do but this methodology is flawed. A native spreadsheet agent with proper tooling will do a much better job understanding the model logic and context than an LLM's chat interface

I'll provide a little more specificity on this, and snippets of an example. For many months people have been talking about a "Cursor moment" in finance, where workflow changes so dramatically that you hit the steep part of an adoption curve. I've been highly skeptical of that, for a few reasons. But the most fundamental reason is the LLM technology just wasn't there. The foundation models simply did not have enough power to interact with Excel spreadsheets in any sort of usable way (despite splashy demos...). Even if you solve the (very hairy) data challenges, 2025-era LLMs just didn't have the power to interact with spreadsheets. So we could sit and talk about a lot of ideas and concepts on how AI could augment institutional investment research. But it was just that, a concept. I have a series of tests I run on new AI models that are capability tests for hedge fund style research workflows. And the easiest is just uploading an existing Excel file to see if the LLM can understand what's going on. If LLMs can't sufficiently read and understand an Excel model, the full stack of AI Excel workflows is just not possible (in my opinion). And a waste of time to try to explore. This didn't work to any sort of impressive degree (Opus 4.6 could do it, but not do it well). Until yesterday, with GPT-5.4 Thinking. Suddenly, I can now get something that is not only modestly useful, but I think will immediately become part of my investment process workflow. I call it "PM Review", or a structured evaluation and push back on a model. I have participated in literally hundreds of these as both analyst and PM. Effectively the analyst builds a model, sends it to the PM, and they walk through it together. The wise, experienced-scarred PM will rip the model apart, push back, and help steer the model to a usable outcome. A great PM will be able to hone in on the two or three key variables that matter and identify aggressive or conservative assumptions. An analyst may be pitching a stock where the core quantitative input is supported by flawed logic. And the PM's job is to try and identify that flawed logic. This workflow, to me, is a key differentiator between good and not good PMs. However this workflow isn't just for PMs; it's for analysts who are trying to evaluate their own work, peer analysts who want to do thoughtful push-back on ideas the team may participate in, our director of research teams who are looking to efficiently evaluate the idea underwriting process. Or PMs for the first cut if they're looking at lots of ideas. The intriguing aspect of augmenting this process with AI is it scales incredibly. And it can run autonomously. Across 300 models I could have a swarm of agents doing automated due diligence on the key drivers, updating those models, feeding those results back to me, and flagging which of my covered ideas have earnings revision potential. This workflow is the "Cursor moment" for public equity research, in my opinion. I'm not saying we're there by any means as data accuracy and the structures required to incorporate internal data are still in progress. But we just took a step forward in the technological capability. I tested this out in GPT-5.4. And while it's not perfect this is the first time I've received anything that's useful back in this test. I'll walk you through a couple of steps to do this on your own. Step 1: brain dump into Claude. I don't know if there's any logic to it or just my own habit but if I'm executing in Chat GPT, I'll meta prompt and Claude and vice versa. I'm not sure where you meta prompt matters all that much for the types of workflows I do but it CERTAINLY matters if you meta prompt vs. raw prompt so don't skip this step. Step 2: take that prompt output, turn it into Markdown, and put that as custom instructions in a GPT project. This is just a workflow efficiency because then I now have a GPT project that I can upload any model into. Step 3: run the prompt. I purposely jacked up my DraftKings model a little bit (and it's a work in progress anyway so do not take any of these estimates as anything I believe). But it produced an exceptionally helpful: 1) Executive Summary 2) Business understanding (explaining how a dollar flows through P&L) 3) Model Evaluation, providing an assessment and sanity check of all of the key inputs 4) Model audit, looking for input consistency, formula integrity, and broken references 5) A road map for incremental due diligence 6) The highest value IR questions I encourage you to check it out for yourself. Will link to the six-page output in the replies.

@modic123 @tryshortcutai Excel is special, because like code, it is the reasoning layer, but unlike code, it is also the output layer. This means formatting is even more important. Opus 4.6 is much better at formatting than GPT 5.4, probably for the same reasons it is also better at frontend ui.

We had early access to GPT 5.4 to benchmark on @tryshortcutai and it is extremely good at Excel Here is a quick walkthrough of two real-world tasks in our internal eval of 50,000+ cells each Watch, because the industry is way beyond the same flashy and meaningless DCF demo

@nvr_stochastic @tryshortcutai marginal differences. higher accuracy, worse formatting for GPT. it's a matter of preference/task type

@phamrich_ @tryshortcutai What about high comparisons?

GPT-5.4 is now available in Shortcut @tryshortcutai

@alexgrama @tryshortcutai settings -> controls

@phamrich_ @tryshortcutai how do we choose the model?

if you don’t say please and thank you to your AI we probably have nothing in common