Nothing unusual to see here, move along.
Joined March 2014
- Tweets62
- Following24
- Followers15
- Likes52
Tracked on the status page now posthogstatus.com/us?inciden…
@posthog It seems like us.i.posthog.com may be having issues? Seeing timeouts and flag compute failure flags in responses. No notes on status page yet.
One of these days, someone is going to craft a supply chain compromise to trigger on and infect a widely adopted package being updated to patch a critical vulnerability. It's the logical way to maximize uptake as people panic to upgrade. That day will be a very bad day.
First you want to have some kind of data capture in place, there are a lot of commodity tools that can do that for you already (sales email logging, call transcription, etc). With that then you basically need something that "reads" all the unstructured data to recover missing records. The hard part is mostly how you do that at very high scale in a cost effective, reliable, accurate way while still applying a level of reasoning and judgement to what is relevant to recover, and then how important that was to the deal in question in context. We've spent a long time working on those problems, with more to go, but what we see already is that many orgs have most of the data the need somewhere but it needs that reconstructive process to be legible.
If you want to scale your software dev without devolving into vibe coded slop, get really familiar with methodical, verifiable risk analysis, mitigations, and compensating controls. Not just for security, but especially for security.
The difference between vibe coding and a software factory boils down to two things:
- automation (work keeps going without a human pressing go)
- really, really good scaled risk management (security risks, behavior defect risks, availability risks)
Jonas Bauer retweeted
Clay just shipped what might be the best version anyone will ever build of the 2025 definition of 'Cursor for GTM'.
This category still won't survive, even for Clay.
What they launched today at Sculpt: an agent harness, a Knowledge Hub for context, and a pitch to work with Clay "the way software engineers work with coding agents." It's exactly what they had to build, and they're probably the only company with any hope of pulling it off.
That background wailing is fifty other 'Cursor for GTM' startups turning back into pumpkins 🎃
But GTM engineers already have coding agents: Claude Code, Codex, Cursor. Clay knows it (Workflows also runs from "the coding agent of your choice via CLI"). The labs won't cede a layer this valuable, and they get to train the horse in harness because they own both.
Clay's own learning example suggests an ICP change after a campaign 'books six meetings.' That's not an eval. Meetings aren't revenue, and the truth behind deals is buried in a stack built for the age of humans.
You can't solve 'what worked' at the orchestration layer.
Jonas Bauer retweeted
There is nothing cuter than @camillericketts, @swyx and Alma reading my book about Dragon, Mina and attribution. I smiled and laughed every time I watched this and now I want a recipe for these pepper buns and a stuffed dragon for @UpsideGTM 🙃
yes, the question is when, not if. The automations will eventually perform better review than humans on all dimensions, at which point it becomes irresponsible to NOT use them. Humans should, for a while, still be involved in the change management approval, just not code review.
Astra/Fable class models don’t need micro managing, and worse, you can actively make them worse if you try to. Learn how to draw the rest of the owl with good delegation up front, and then getting out of the way.
If I need 'forced creativity' I often tell my agent to read en.wikipedia.org/wiki/Specia… first, and then perform its task. Similar effect, it injects random context into the model before it proceeds.
Obviously with the volume of code production from the factory, you're going to need one of these to submit a PR to Github.
Does the DuckLabs team joining AWS mean we're going to get some kind of DuckRDS? I would love to see some sort of managed AuroraDuck fusion for hybrid workloads