HN user
carsonkahn.com | carson@carsonkahn.com
Our team at Agentforce recently open-sourced our solution to this and we've gotten very valuable feedback -- would love to hear from more of you about it: https://github.com/salesforce/agentscript
Ditto Ethan's point -- and hundreds of customers tell us it works very well. We'd value more feedback from this community, not just the Salesforce/Agentforce customer base!
The headline misses the most impressive part: ChatGPT one-shotted the problem. No turns, no retries, no mid-thinking steering from the user. One-shotting a problem like this would have been nearly unthinkable in 2025.
This (sort of) is already a paradigm: https://en.m.wikipedia.org/wiki/Probabilistic_programming
I’m building software to augment human cognition.
In particular:
To help solve forecasting & planning problems too hard to hold in your head, I’m converting natural-language formulations of constrained optimization problems into (back)solvable mathematical programs, whose candidate solutions are “scenarios” in a multi-dimensional “scenario landscape” that can be pivoted, filtered, or otherwise interrogated by an LLM equipped with analytical tools:
- 5 minute demo: https://youtu.be/-QdqiLp_9nY
- Details: https://spindle.ai
Eager to connect with anyone interested in similarly neurosymbolic “tools for thought”: carson@carsonkahn.com | +1 (303) 808-5874
Very f’ing cool (esp. optimistic about repo-level codebase completion) – but just like many other results that DeepSeek reports, their preprint leaves me with more questions than they’ve given answers, unless I’ve misunderstood multiple pieces of it (which of course is possible):
—They report a 9.0× speedup in forward pass and 6.0× in backward pass… Why the heck would the backward pass be so much slower? Is it their gating mechanisms needing extra computation in backward passes? Gradient accumulation or KV-cache updates bottlenecking the speedup? FlashAttention (or at least FlashAttention-2) gives a near-equal forward-backward efficiency… They claim it’s tuned for FA2-style blockwise layouts, so which of their (competing) claims is wrong?
—Does NSA actually learn useful sparsity, or just get lucky with pretraining? How much of the performance gain comes from pretrained sparsity patterns vs. sparsity inherent to the attention? Even though they themselves say “applying sparsity post-hoc forces models to deviate from their pretrained optimization trajectory… As demonstrated by Chen et al. (2024), [sic] top 20% attention can only cover 70% of the total attention scores, rendering structures like retrieval heads in pretrained models vulnerable to pruning during inference” — yet their ablation isn’t strong enough to tell. A stronger ablation would include (1) a Full Attention → NSA transition test to measure whether NSA can be applied post-hoc without degradation, (2) a visualization of learned sparsity patterns over training epochs, and (3) a test where sparsity constraints are randomly assigned to see if NSA actually finds useful structures or just adapts to imposed ones.
—Training transformers with sparse attention is historically unstable — early MoEs like Switch-Transformer (which use expert gating-like mechanisms just like this one) were famous specifically for their collapse issues. How does NSA prevent mode collapse in early training — or really, how do we know it’s not just going to collapse different (i.e. more common) initialization schemes? If their technique doesn’t have an explicit mechanism for counteracting sparse expert underutilization, then it’s just as vulnerable to collapse as (e.g.) Switch-Transformer — but worse, since sparsity here isn’t just a gating function, it’s the core of the entire attention mechanism…
I’ve been working on a glossary of “Notions of Equality & Sameness” every few months for the last ~8 years (and looking for collaborators: carson@carsonkahn.com) - surprised not to have had this one in my list already!
I'm working actively on this area – anyone interested in helping should please get in touch! carson@spindle.app
Anyone else find it interesting/odd that their UA string reports a WebKit (Safari?) environment rather than Blink/Chrome?
Feature request: Index arbitrary content/history across browsers and allow it to be re-opened (my main use case for your pseudo-competitor Rewind.ai).
@daly I’m working on something relevant here and would love to discuss; email me: carson@carsonkahn.com
What about Dromo? https://www.dromo.io
Your website pitches this product SO well. Kudos.
As a data scientist/machine learning engineer, I read "EDA" as "Exploratory Data Analysis" [0] and was amused (if let down) when I clicked through.
Luckily, that’s not my application area if these ideas right now. :) But it might indeed fail for some of the same structural regularities!
I’m working on a commercial side project along these lines — if anyone is interested in helping or collaborating, shoot me an email!
I'm not looking for a cofounder, but I love helping people who are: carsonkahn.com/who
I’d like to chat more about this (including perhaps hiring you for a related project). Mind emailing me? carson@carsonkahn.com
Just realized you may not have access to texting. Email me instead if you need to. We’re going to make this all work out for you.
My name is Carson (carsonkahn.com/who) and I want to do whatever I possibly can to help you. Text me and let’s see what we can make happen for you: +1 (303) 808-5874. It is all going to be OK — not because of some mystical, arbitrary self-assurance, but because we’re going to put our brains and hearts into it. Don’t “throw yourself into a river”, please; text me. It will be okay.
Location: SF Remote: No preference Willing to relocate: Yes Résumé/CV: carsonkahn.com/who Email: carson@carsonkahn.com
I’m a moderately successful serial founder, most recently CTO promoted by F100 Board members to CEO.
Very open-minded about opps, given unique background.
IMHO, it’s an antipattern in almost every way. Being “common” doesn’t make it “smart”. The preprocessing argument is invalid for a lake, and the upfront “integration” time savings are cannibalized later by maintenance and risk overhead vs. proper service integrations. And everything you mentioned aside, there’s security and DPP risk. Lakes should not be used for collaboration between systems - that’s what lakeshore data marts are for, at worst, or real service-to-service APIs, ideally. (We have a good word for this in German: Datensparsamkeit. See also https://martinfowler.com/bliki/Datensparsamkeit.html). The people and services regularly using the lake should be data scientists and analytics folks, probably not prod services.
The right tool for the job