iiinova — where your digital self keeps evolving

Eric Kennedy @ekeric13.bsky.social · 01/09/2026

How I Plan an Engineering Project End to End Caveat: This is how I operate at a big tech company. This would look different at a startup and vastly different if I was just building alone. At a large company there is definitely a coordination tax you need to pay, and that seeps into how you build.

120

Eric Kennedy @ekeric13.bsky.social · 06/08/2026

feels like the most successful neural network arch is designed 1:1 with real world hardware limitations. most obvious example is the transformer allowing for parallelism that lstm did not. but a more precise example is moe maps to gpu rack layout.

000

Eric Kennedy @ekeric13.bsky.social · 04/08/2026

AI has let people develop systems that cannot be managed or reasoned about without the assistance of AI, creating a lock-in effect and a dependency on the models. Being able to utilize a complex system can be a competitive advantage, but 1/n

110

Eric Kennedy @ekeric13.bsky.social · 22/07/2026

The amazing intelligence of these models combined with the startling limitation of the context window is going to lead to a memento lie scenario where they leave a bunch of random files around to help construct Skynet

011

Eric Kennedy @ekeric13.bsky.social · 21/07/2026

ML engineers understand what it means to be GPU poor. It limits what experiments you can run, how fast you can iterate, and what work you can realistically take on. Software engineers are about to understand what it means to be token poor.

100

Eric Kennedy @ekeric13.bsky.social · 17/07/2026

fable = vibe coder model sol = read the code model

010

Reposted by Eric Kennedy

P(aul) Frazee @pfrazee.com · 16/07/2026

Prompt injecting your coworker into ordering lunch via PR review

1433020

Eric Kennedy @ekeric13.bsky.social · 14/07/2026

The reason why I think bad AI writing bugs people so much - it is genuinely disrespectful to ask someone to put effort into reading something you couldn’t be bothered to put an ounce of effort into writing...

110

Eric Kennedy @ekeric13.bsky.social · 08/07/2026

A software engineering truism that no longer makes sense after AI codegen: "Never do a full rewrite." Historically, rewrites were always a terrible idea. What would happen is the team feels like they have learned so much more about the business since the original system

110

Eric Kennedy @ekeric13.bsky.social · 06/02/2025

i would argue we have had AGI since early 2024 and it was generally accessible It doesn't seem like that will be the case as we move towards ASI - $200 a month price limit for openai, rate limits for claude, etc. It does seem like super human intelligence will only be accessible to those already

110

Eric Kennedy @ekeric13.bsky.social · 01/02/2025

It’s been months since I heard about RAG. I figured large context windows would consume it but surprised about how fast. Vector databases still have a space but in terms of a tacked on framework seems like agents have replaced RAG as the cherry on top of the llm cake

000

Eric Kennedy @ekeric13.bsky.social · 27/01/2025

being able to see the stream of thought for reasoning model is super useful for getting the query right in a follow-up

000

Eric Kennedy @ekeric13.bsky.social · 27/01/2025

the strength of the deepseek models is a big win for all llm app developers who have been building local first non-http apps. i think that space will get really interesting.

000

Eric Kennedy @ekeric13.bsky.social · 25/01/2025

openai's task feature very hit or miss. using it to give me fun weekend things to do with family and every single one of them is a hallucination. it even has links it cites (they are all eventbrite for some reason) and the underlying website doesn't contain the activity/event

000

Eric Kennedy @ekeric13.bsky.social · 23/01/2025

so the deepseek guys are definitely heros and trailblazers. and of course the irony isn't lost on anyone that china leading the open source charge. but is it china or is it deepseek? like is tencent and bytedance open sourcing their key algorithms to the same extent or is it like openai/anthropic?

010

Eric Kennedy @ekeric13.bsky.social · 22/01/2025

since 2022 there has been a bit of a horse race between llm models and image gen models. there is more utility in llms so of course more chatter around them but it seems like the hype around image gen has really died down.

000

Eric Kennedy @ekeric13.bsky.social · 18/01/2025

uv seriously bumps up python in terms of the language rankings. is crazy how bad python's package management ecosystem has been. it's weird that poetry wasn't the answer - it has everything you want in theory but it always feels like a hassel

000

Eric Kennedy @ekeric13.bsky.social · 17/01/2025

it is so much easier to code now - the cost of code has gone way down. there is way less of a premium on talented coders. llms still don't understand code from an app perspective though. it rarely recommends using monitoring/telemetry/analytics in a way that scales. so there is still value in...

230