Alex Kemper

Having fun with autoresearch. More than doubled a search agent's primary metric by running this in the background while I was on calls yesterday afternoon

We made long-context opt-in back in December for this very reason. I'd guess half of our support cases back then were complaints about quality and cost over runs, and long-context use was the cause of virtually all of them

the year is 2027. garry tan has just crossed 1b lines of code per day. water to 3 rural californian towns were diverted in order to cool his locally ran LLMs. riots erupt, and protesters demand answers to one, single question: "what is he building?"

Really good write up of what FDE work looks like today. The point about building a shared set of dashboards and evals that serve as a mutual source of truth especially resonated. We've seen the same at Amp — doors unlock and the project accelerates when you have a nice graph that shows progress over time. A picture is worth a thousand words, as they say!

Talking about “way more software will be created in the future” is like saying “when everyone learned to read and write, there were way more books”. Sure, it’s true, but the arguably bigger and harder-to-forecast impact was from all the one-off reading and writing.

Ramp says companies that spend a lot on AI are growing faster I'm skeptical this is anything but correlation In our experience, applied AI improves productivity via back office automation Connection to top-line growth is tenuous (for now) It's an exercise in margin expansion rather than growth acceleration That said, fast-growing, dynamic companies investing in AI today will overtake their competitors in markets where moats are weak or shifting, which should eventually flow through to growth

Here's some current Amp meta.

@Skulledwedge Bump

Listen Lassy, you could take your signing bonus and have $10 million in cash right now. But that’s short term thinking. Once Big Daddy Trump gets the regulations fixed Sable stock will be worth 50, 60, maybe even 100x that. We’ll be sipping Coffee for Wellness on our yachts in the Persian Gulf by the end of the year. What do you say? Let’s do this deal

Great post. Especially curious about option 5 here

Because of the wealth generated by Carnegie Steel and Standard Oil, Pittsburgh and Cleveland have some of the best architecture in the country from that era Even the smaller cities in that region like Akron, Youngstown, etc that are now totally bombed out have historic neighborhoods that rival those in the NYC area

Lobby of the hotel we stayed at in Cleveland. 12-story department store built in 1910 that was later converted into a residential building.

@dboskovic Yep, the entire system and processes around software delivery are going to have to be redesigned Lots of big enterprises frustrated by "unrealized gains" of coding agents have delivery timelines of hours-to-days for code changes!

@AJKemps what is phenomenal to realize is that if you're shipping 1000 PRs a day w/ agents that can do rounds of review in 30s, the bottleneck is not AI, it's very quickly the things we've accepted as reasonable latency and use of resources (refetching a 3gb cache on every shard)

Cool to see how people are cutting CI time from 5 min to 15 seconds to relieve one of the big software delivery bottlenecks in our post-agentic world

coming right up but TLDR Cut out all setup overhead (prev 30s, now about 1s) - ovh amd epic turin box with 128 cores and 256gb ram (1k/mo) - golden image of main with all cache loaded - zfs for instant copy of golden image (this is magic) - git fetch all every second for local mirror - golden image of database so only last migration runs (as Postgres template) - turbo cache locally For actual suites - much higher sharding since now no overhead to each shard - use @bunjavascript tests where possible to avoid typescript compilation - incremental typechecking with local cache Bypassing GitHub actions in favor of custom check suites - a few seconds of queue time saved - no action minutes billed (we hit 36k minutes in 3 days) For preview apps - JiT full stack preview apps (not deployed on each commit) - 2-3s cold start on any commit sha to a fully deployed full stack preview app - zfs clone of golden firecracker vm and then check out latest commit etc

@dboskovic Would love to see a little blog post on how you did this

We refactored our CI this week from 5 minutes to about 15 seconds on a single bare metal machine. We've been shipping so many PRs that CI time was becoming both a cost center and bottleneck. 1/10th the cost 1/10th the time 100x the PRs

A few thoughts about this Block news: - the vast majority of companies like Block could reduce headcount by 50% today and it would not materially degrade their ability to operate and serve customers - this cut is not about actual realized efficiencies from applied AI today - rather, most businesses like Block have suspected they could make similar cuts for years - execs inside these companies will often say something like "we have 3x the headcount today v 2020, but we don't get anything done faster or cheaper" - with the progress of applied AI over the past few months, execs like Dorsey look at the p(success) for massive cuts like this and see it move from 60% last year to 95% now - so this is really just about "we've wanted to make these cuts for a while anyway and now we have very high confidence the cuts will work out because of the early promise we're seeing from applied AI inside the org today" - there's a very real mimetic contagion risk here; like we said earlier, virtually all companies like Block have been tempted to make similar cuts over the previous year or two (with varying degrees of seriousness) - I expect we'll see more of this happening over the coming months for this reason - we'll likely see a different class of reorgs and layoffs late this year or early next, and that'll be about actual realized gains from applied AI (distinct from the Block layoff which is more like a bet about the promise of these tools, and getting their house in order, but it's a huge bet just by the proportion of their employees impacted) P.S. — my heart goes out to everyone impacted by this cut. I've been through it myself, and here's to hoping the best work of their careers are ahead of them.

@howietl @jeff_weinstein Nice. Would love to be an early tester!

I've been personally burning through billions of tokens a week for the past few months as a builder. Today I'm excited to announce Hyperagent, by Airtable. An agents platform where every session gets its own isolated, full computing environment in the cloud — no Mac Mini required. Real browser, code execution, image/video generation, data warehouse access, hundreds of integrations, and the ability to learn any new API as a skill. Deep domain expertise through skill learning. Teach the agent how your firm evaluates startups or how your team runs due diligence — now anyone on the team gets output that reflects your actual methodology, not a generic template. One-click deployment into Slack as intelligent coworkers. These aren't bots that wait to be @mentioned — they follow conversations, understand context, and act when relevant. And a command center to oversee and continuously improve your entire fleet of agents at scale. We're onboarding early users now. hyperagent.com

Tempted to write a blog post about why the better analogy for Anthropic, OpenAI, et al is actually the US steel industry post PRC joining the WTO, and not cloud hyperscalers of the last decade Cournot Competition and Nash Equilibria only model economically-rational participants

bold new statement by ampcode

@benhylak @raindrop_ai This looks great! Was just working on a home-grown "help me search through agent traces in natural language" tool this morning

we’re excited to announce trajectory explorer: the first sane way to navigate agent traces. every decision your agent made is now searchable in seconds only in @raindrop_ai