Dan Cleary

"When we (@intercom) started this pursuit, I honestly didn't think it was going to be possible to be better than frontier models. I thought maybe we can get parity on average. But can we actually beat them?" They did. @PedroTabacof, principal ML scientist at Intercom on how Apex replaced Claude as the core model for Fin, across 2 million customer service conversations a week. Full episode out now. Link below. Amazing work going on over at Intercom @eoghan @fergal_reid Which AI app company is next to launch their own model? @harvey ? @Replit or @Lovable ?

Opus 4.7 vs GPT 5.5 in their respective harnesses Opus with the W

more here: danjcleary.substack.com/p/are-there-an…

Moats never existed, and ChatGPT's memory getting pwnd with a single prompt is just another example of this (and a killer move from Anthropic, matched by Google a few weeks after) More thoughts on moats (or the lack of) in the age of AI below

Convalytics(.dev) is up and running. Easy web + product analytics for @convex apps. Your agent can install with a single prompt and set up custom events for you

"When we (@intercom) started this pursuit, I honestly didn't think it was going to be possible to be better than frontier models. I thought maybe we can get parity on average. But can we actually beat them?" They did. @PedroTabacof, principal ML scientist at Intercom on how Apex replaced Claude as the core model for Fin, across 2 million customer service conversations a week. Full episode out now. Link below. Amazing work going on over at Intercom @eoghan @fergal_reid Which AI app company is next to launch their own model? @harvey ? @Replit or @Lovable ?

Data: slop-bench.vercel.app Code: github.com/Dan-Cleary/slo…

SlopBench: Opus 4.7 did worse (more slop) then Opus 4.6 Code + data on Github and below

Full vid: youtu.be/sajfGkwc4D4

Vibe coding breakdown: Opus 4.7 vs GPT 5.4 vs Gemini 3.1 TL;DR GPT 5.4 Opus 4.7 Sonnet 4.6 Gemini 3.1

Opus 4.7 fav word is malware, idk what is in there but every single edit or file read seems to have to be checked for malware first

Convalytics Convex component is live 🚀

Opus 4.7 benchmarks sans Mythos

Been a consumer of @convex components for awhile, stocked to be publishing my first for convalytics.dev

Youtube:youtube.com/watch?v=nOQY10…

A year ago I was skeptical that companies could train their own models to beat frontier models. @eoghan and the team at Intercom proved me wrong. And this conversation explained exactly how they did it. @PedroTabacof is a Principal ML Scientist at Intercom working on their custom model Apex. -It beats GPT-5.4. -It beats Opus 4.5. -It hallucinates 65% less than Sonnet. And 100% of Fin's traffic now runs on it. We got into: -How they actually post-train a model from scratch -Why evals are 90% of the work -The state of open source models If you care about AI products, vertical models, or where this is all going , this is a must watch. Link below.

Did a whole deep dive here: danjcleary.substack.com/p/is-vertical-…

This didn't get enough attention Intercom built a model that outperforms Opus + GPT 5.4. "As of last week, ~100% of all (English language, chat and email) customer conversations are now running on Apex." Vertical AI is just getting started

@stopachka Love this direction for backends, sounds Convex, so did a quick test for gang x.com/DanJCleary/sta…

After 4 years, we’re announcing Instant 1.0. Instant is the best backend for AI-coded apps. Let us tell you why.

@benroff_ @stopachka x.com/DanJCleary/sta…

@stopachka convex vs instant blog when?