Phenomenal takes in here “But here's an interesting question - why has nobody truly been paying close attention to matching capabilities to the unlock? Where is this discourse happening? The root cause of all the shallowness in thinking is that the wrong category was crowned far too early: the AI-native ERP. That is fundamentally the wrong abstraction for what AI changes in finance.”

RT @ashugarg: Model capabilities have unlocked every category Heres why finance is having its inflection moment

Finance is entering its Legora moment

Really like this framing @JayaGup10. And thank you for the shoutout to @maximor_ai! One nuance I’d add from what we’re seeing at @maximor_ai is that these two things can be true at the same time. For a lot of enterprise work, @JayaGup10 is right that there is an intelligence threshold. Once you’ve solved the same class of problem enough times, more of that judgment should get encoded into policies, checks and deterministic workflows instead of being re-reasoned from scratch every time. But in accounting, we’re still hill-climbing on capability itself. There is a huge amount of genuinely hard work left: understanding non-standard contracts, reconstructing judgment from years of workpapers, reasoning across multiple systems, handling exceptions across entities and currencies, and getting to an answer that a controller, CFO and ultimately an auditor are willing to trust. We still benefit enormously every time frontier models get better. That is Arc 1 for us: get increasingly complex accounting work right. The part I find interesting is that we’re already starting to see the next thing, even while there’s still a lot of Arc 1 left to solve. Take a usage-based software company. To account for revenue correctly, @maximor_ai may need to understand the contract a salesperson negotiated, the committed usage tier, actual product consumption, the invoices that went out, collections, and what ultimately hit the financial statements. Once you understand that whole chain, you can ask a much more valuable question than “what revenue should we recognize?” You can see that customers are consistently being sold the wrong consumption tier, that actual usage is diverging from what was contracted, and that the pricing or packaging decision itself is producing a worse financial outcome. That recommendation only becomes possible because you did the accounting work deeply enough to understand how the business actually behaves. That’s the second arc for me. Arc 1 is making increasingly complex accounting work autonomous. Arc 2 is using the understanding you get from doing that work to help the business make better decisions in the first place.

I love how the one specific analogy @thejessezhang makes here on what is still unsolved today is related to accounting: “If you’re building an AI agent for accounting in 2026, there is no established workflow, because literally nobody has ever used one of these. Nobody knows what the user journey looks like, not you, and importantly, not your customer either. They can’t tell you what they want, because the thing they’d want doesn’t have a shape yet.” The reality is all we’ve been doing at @maximor_ai over the past year and a half is figuring out how to embed our engineers into large enterprises and “excrete” accounting and finance workflows that generalizes and allows our customer set to span across verticals, sizes and business models. We’ve spent a lot of time thinking about where the boundary of “compile time” code and “inference time” code lies and how our Policy layer allows for the right level of abstraction. What we’ve seen is that UX design has become so much more complex than before, to the point where the more we automate the work, the user persona itself changes as a function of time. What took multiple Staff Accountants to prepare is now just reviewed by the Controller. Our rationale is simple. The best product with the best design wins. We have FDEs and we’ve made their goal a simple 2 step process: Get to the customer outcome as fast as possible. Then build repeatable UX by creating the right product and platform primitives . Get stuck in the first step and we don’t even count those as real recurring revenue internally.

Fundamentals of programming are as important as ever. In fact, I’d argue they’re more important than ever because spotting the one mistake (the needle), in a 5000 line AI generated commit/PR (the haystack), is almost impossible if you don’t know what you’re looking for. If you can think in terms of specs - product and engineering, deliberately documenting the relative importance of tradeoffs, architecture, edge cases and invariants, you’re much more likely to get your coding harnesses to the “ideal state” of your code and do it much faster in terms of time and $$s spent. I’ve repeatedly seen how hard it is for engineers who started programming in the post GPT era to truly be disciplined about these “basic” practices simply because they were never forced to develop the same feedback loops. I remember when I wrote my first PR at Microsoft back in 2016, the heart of the problem I was trying to solve was conserving ports for the DNS forwarder application we’d built to convert DNS requests and responses from physical to virtual machines and vice versa. And the first thing I learnt by undergoing quite a lot of pain was that the unintended consequences of an otherwise reasonable optimization could be enormous, and it was important to think through what I might be silently trading off in every decision path. I vaguely remember the first version of my solution being perfectly reasonable for the average workload but inadvertently throttling requests for certain VMs which had very bursty DNS request patterns, which ended up needing a very nontrivial workaround. I had optimized one constraint without sufficiently modeling the shape of the workload and the other constraints the system needed to preserve. No coding agent today is getting to the ideal solution to a sufficiently complicated problem in one shot. But the more you can ground it in edge cases and invariants and continue to refine the relative importance of tradeoffs, the more you narrow the search space enough to get there really fast. Tokens consumed also matter. It’s not just a question of cost, sure that’s a factor if you just spam Fable for everything, but it’s also a way to control the steps the agent takes. Something as simple as having the agent check some global invariants along with local invariants can be the difference between converging and spending a few more hours going in circles. In finance and accounting we’ve seen repeatedly that local decisions sometimes make sense but taken in global context it’s much easier to see why they’re wrong. Reconciliations are classic examples. The structure of the problem is almost exactly like a maximum-weight bipartite matching problem, but when certain global P&L and balance sheet constrains are added, a lot of locally plausible paths become obviously wrong. Tl;Dr: more than ever it’s time to be your most curious self with the need to understand things deeply. Otherwise you’re just going to push increasingly well-hidden slop.

What is particularly impressive about this piece from @JayaGup10 is that it hammers a very particular point - the best way to analyze the impact of a decision is to simply chase the $$s! When a set of companies do poorly because of a ban of the best open source models, what second order effects do they cause? Today’s reality is that the AI market expects near perfect execution to justify everything from valuation to financing. There is a chance that this doesn’t work out as expected because of fairly mundane reasons like datacenter buildouts taking 20% longer than the assumed timeline. But banning the best open source models would be the biggest own goal ever - this simply will not be a contained event. It will ricochet and destroy everything in its path. In fact this alone may do more to harm OpenAI and Anthropic’s valuation with all the downstream that comes with it than any mal-execution from them!!

Distillation is quietly becoming one of the defining questions in AI. Self-distillation, agents, RL, IP - so excited for this conversation - exactly the kind our ecosystem needs as we go from “frontier closed source models solves every meaningful problem” to creating value for every person and business in the world. @JayaGup10 - brilliant idea!

Hosting America’s First Distillation Summit: With distillation fast becoming one of the most consequential and contested topics in AI, we are convening the best minds for a summit. Sneakpeak of the agenda includes: - Is Distillation Enough? Architecture, RL, and the Ingredients Beyond Copying - The Mechanics of On-Policy Distillation - Adversarial Distillation, IP & National Security - Open-Weight vs. Closed Frontier Labs - Scaling to Agents & Long-Horizon Tasks - Self-distillation and self-improvement - Ethics of distillation Excited for researchers, government officials, founders, and executives to convene on this! Of course the best ideas, like the best whiskey, get better when you reduce them to what matters :) And yes, there will be whiskey

It was so much fun having @deedydas over in our office with the @maximor_ai team on Friday. What was supposed to be an hour turned into a much longer, deeper discussion on Glean’s evolution, AI product strategy, engineering culture, value capture, and where the AI ecosystem is headed. Four ideas that stuck with me: • “Lightning in a bottle” is a real startup category. • The difference between a $10B and $100B company is often whether the underlying problem is bounded or unbounded. • AI apps should be judged less by model quality and more by how much workflow they own. • Glean’s superpower wasn’t search—it was reinventing itself for the LLM era instead of protecting its existing product. Thanks again for the incredibly candid conversation!

Some good best practices here on AI token cost optimization. None of these happens though without a deep understanding of the underlying work being done in a non-abstract way. The ultimate implication is that a layer between the work itself and the underlying intelligence needs to deeply understand your workflows, context, and business process. Now, each individual company doing this on their own is unlikely to be effective at scale, so as a consequence, this is effectively the playbook for any applied AI company right now. By evaling the models for the applied use cases, deeply understanding the domain, having tuned UX and features for the use case, and having the ability to support adoption and change (via FDEs), allow this layer to add a ton of value. And as a result, enterprises get higher ROI because you actually can get *more* intelligence per dollar by having optimal architecture and workflows. There will be many horizontal and vertical versions of this approach. Huge opportunity right now.

Uff.. another banger! Here's another macro shift happening: If intelligence is a priced resource, then the CFO becomes the most important person at the company – the one allocating spend & analyzing ROI across the only 4 inputs that matter: customers, employees, vendors, and now.. tokens. That allocation problem is unsolved. We've already figured it out for all of finance. For 20 years IT/engineering decided what companies bought. AI hands that power to the CFO. It's the CFO's turn to take it. We're arming them. @maximor_ai 🫡

Great read!! In 2026, the CFO has become the single most important decision-maker in enterprises. And @maximor_ai being the Clearinghouse (as @jaminball put it) for the Office of the CFO connecting $$s to downstream systems that use them to produce results has made us best positioned to help CFOs allocate tokens effectively across functions.

We’re extremely bullish on non-linear thinkers within @maximor_ai. The ones who can connect the dots between theories and philosophies to actually implementing them rapidly. Karp’s bullishness on neurodivergence has stuck with me - it looks like increasingly the alpha in building companies is going to be found in those who can connect dots that just are almost impossible to see unless you allow your mind to wander in these higher dimensional spaces. Our best engineers have increasingly been “principled dreamers” - ones constantly oscillating between risk and principle and are willing to stake their reputation on uncharted territories.