Amplify Partners

Truly wild that prompt engineering is beating fine-tuning, RL, distillation when it comes to model adaption. Also wild that twice as many AI engineers are using RAG than fine-tuning. The top 3 techniques for adapting model behavior actually don’t touch the original model weights at all. And there’s a huge drop off when you look at techniques that do touch model weights. A few reasons for this: - There’s a skill gap → adapting model weights is much harder than other techniques - High quality synthetic data generation is very hard at scale - There are few good out-of-the-box solutions for fine-tuning Ofc the sample of participants (AI engineers at @aiDotEngineer World Fair) plays a role too, as they likely have less experience with hands-on ML techniques, but it’s still an important subset of people using AI models. Really interesting report from @AmplifyPartners.

One of the (now) most obvious fallacies that's underpinned the AI hype train over the last few years was believing that what mattered most was frontier intelligence (and the gap between frontier and all other models) vs. the baseline intelligence of the ecosystem. As baseline model capability continues to improve and more models approach ~Opus 4.5-level intelligence, users and enterprises will make rational economic decisions about which model to use for each workload. Once we've crossed the "powerful enough" threshold, cost, latency, privacy, specialization, etc. become the deciding factors. Further, we've learned that 1) harness engineering often matters as much as the underlying model and 2) enterprises have every incentive to own their own "learning stack." All of this is reinforced and underpinned by the fact that switching costs between models continue to stay incredibly low. The idea that 1 or 2 foundation models would eat 80-90% of workloads now seems implausible. Every enterprise will route different workloads to different models based on economics and workload characteristics. Net net, we're much more likely to see dozens or even hundreds of models inside every company. That's a much harder world for the big labs. Frontier intelligence will still obviously matter, but we're seeing rapid commoditization for many workloads; you don't need a 300 IQ model to write React features, service a support ticket, write a legal brief, or automate marketing. It's no coincidence every major lab is racing up into vertical applications. On the flip side, to the point below, this creates a much broader ecosystem where infra providers, tooling vendors and app creators stand to massively benefit. Exciting times ahead!

It's been three years since we released Sim TigerBeetle! sim.tigerbeetle.com

AI search is actually five separate problems masquerading as one feature. Decision to search, tool calling, SERP APIs, query generation, citations. Each one is hard. amplifypartners.com/blog-posts/how…

Founders cut great candidates all the time for missing “nice to haves” while they should have crushed the “must haves”. Before you start interviewing, distinguish what actually matters! @NatashaKatoni’s full post: amplifypartners.com/blog-posts/the…

Very interesting results from this year's AI engineering survey, in particular that a full 25% of respondents have agents in production that have write access without humans in the loop (!)

Damn these @_hex_tech charts are slick 💅

lots of gems in this. amazing stuff from @barrnanas. a few that hit close to home :) - “ open-weight models aren’t replacing closed models. they’re complementing them” -> hybrid inference ftw 🙌🏽 - “cost is now a first-class constraint” -> intelligence per watt matters ☺️

@barrnanas's AIE roundup is an annual tradition great thread! fascinating that evals are *still* the #1 priority for teams building, 3 years in a row

The trend here is the interesting part. In a year I expect this chart to be quite different (including app layer models trained from open-weight base)

open-weight models complement frontier models ie the goal is not to redistribute the pie, but to make it bigger

1/ Last week at @aiDotEngineer, I presented the 2026 AI Engineering Survey: 1,000+ AI engineers on model selection, build vs. buy, who’s shipping with AI, and (of course) whether GPUs are going to space. This year, we ran it with @NotionHQ and @vercel. Some highlights 🧵

2026 AI engineering survey results 🌝👇🌶️

@polymorph3us @BenjDicken bet amplifypartners.com/blog-posts/the…

@BenjDicken didn't @AmplifyPartners write about a team in germany doing precisely that?

I'm curious what it would take a talented engineer to build Postgres from scratch in 2026. Not a slopfork. Written in Rust, from the ground up, and fixing the mistakes of the original design (process per connection, table bloat, schema catalog). Who wants to burn some tokens?

For RL, a training step can take 90 seconds while generation takes 10+ minutes. With GPU prices what they are, this idle time is highly undesirable. From embracing staleness to PipelineRL’s distributed systems approach, here is how labs are maximizing generation efficiency: amplifypartners.com/blog-posts/max…

Docker images take minutes to pull. @modal built a content-addressed filesystem that gets container cold starts down to seconds — even with Torch. How they did it: amplifypartners.com/blog-posts/how…

Until next year, @aiDotEngineer

The biggest hiring mistake at seed stage: hiring for where you think you’ll be in five years instead of where you are today. A great engineer is only great if they’re the right engineer for your current stage. How to build your hiring plan: amplifypartners.com/blog-posts/how…

Like @lennypruss I have become pi-pilled and think its approach –– tight, scoped primitives with extremely flexible extensibility built on top –– is going to be a powerful approach across a bunch of domains of software.

“Features are a liability for agents. Every feature expands the decision space an agent must reason about. More features mean more edge cases, more ambiguity, and more failure modes.” so good, thank you for writing this @lennypruss