Joe Muoio

Full stack freelance software developer and consultant building AI integrations and Cloud services. I like games and food.

[Joe Muoio @JoeMoCode]

@thorstenball I use PRs on solo projects now for review/checks. No it's not so I read every line of code and "catch things opus didn't"...it's to catch things models CANNOT.

For instance, a model cannot know if your real intent was met; it wont know the long term view of architecture, etc

[Joe Muoio @JoeMoCode]

@LKRBuilds @omarvvvr "kids these days..."

[Joe Muoio @JoeMoCode]

@omarvvvr Let's me stay in the CLI. One window, no distraction.

I use herdr as my multiplexer of choice.

Also gives same experience if I am on a remote server or local.

[Joe Muoio @JoeMoCode]

@habibislop @realZahedSaifi @gmkurtzer While I agree with you, look around and see the number of people who truly believe LLMs think...

[Joe Muoio @JoeMoCode]

@realZahedSaifi @gmkurtzer Nice, succinct way to put this!

[Joe Muoio @JoeMoCode]

@zipeng365 I originally used this for different LLMs to Code/Do vs plan/orchestrate. I wanted to save expensive, high context window models for large context work and use composer 2.5 for actually doing the work much cheaper. I still use the pattern even with (mostly) using a single model

[Joe Muoio @JoeMoCode]

@zipeng365 yeah, the kids compact when task is large. You definitely might spend more this way in certain cases, especially when the instructions to do the thing are less than the output tokens needed lol.

[Joe Muoio @JoeMoCode]

@allenholub Big agree. You're easily compressing many times more decisions into a smaller timeline right now developing with AI. Even if you are managing a few agents

I am lucky to learn about context switching burnout pre-ai. Truly, the answer is to be mindful of your needs

[Joe Muoio @JoeMoCode]

@plainionist compaction is inherently expensive since you blow the cache. So my opinion is to just start a new session instead. And try to scope work to a single session, using worktrees to manage parallel work. Stay in the "smart" zone

[Joe Muoio @JoeMoCode]

@plainionist I avoid compaction. Rather than compact before a slice, I use subagents orchestrated from a parent session which handles each subagent feature, merges the work, then verifies based on success criteria defined in a plan for the work.

Might save even more tokens ;)

[Joe Muoio @JoeMoCode]

@gmkurtzer I mean, you CAN reinforce weights during inference...But that's not what most deployments of models do (or so, I hope!)

This does not take away from your main point since that's definitely not how memory works in mammal brains :)

[Joe Muoio @JoeMoCode]

@ALEngineered Why would it?

[Joe Muoio @JoeMoCode]

@johncrickett @fjzeit Don't threaten me with a good time

[Grady Booch @Grady_Booch]

In the face of generative AI coding agents, the role of human-led architecture becomes even more important in the crafting of software-intensive systems.

All architecture is design, but not all design is architecture. Architecture represents the set of significant design decisions that shape the form and function of a system, where significant as measured by cost of change.

As such, remember that coding agents are reasonably good at transforming prompts into code, but they are not particularly good at making decisions that balance the static and dynamic forces on a system in order to yield a reasonably optimal solution: this is where human judgment, human understanding, human knowledge, and human experience shine.

[Joe Muoio @JoeMoCode]

@infinitehumanai You might get better results with "worse" models?

Less built-in world knowledge gives you less to fight against. 🤷

Been a fan of deepseek v4 flash for general purpose work. Needs the right grounding, of course!

[Joe Muoio @JoeMoCode]

Bingo!

This is a huge part of the latest string of doomerism calling for regulation. Regulation that reinforces the abdication of human responsibility. Not good.

Also, they are running out of money...

[Carl Quintanilla @carlquintanilla]

“.. The greatest danger facing society isn’t that software will awaken and overthrow its human masters. It is that we will allow the creators of the software to abdicate human responsibility for the systems they choose to build ..”

@WSJopinion
https://t.co/zXHA3U4h5F

[Joe Muoio @JoeMoCode]

@SenSanders Stop spreading the lies of AI companies. They are going no where and have been negligent in their deployment of their tech and need money. There are no AIs that function outside of our control. Total lie.

And if I am wrong, prove it. Subpoena them. Get real proof.

[Joe Muoio @JoeMoCode]

@mrdoob Appreciate you, MrDoob! Haters gonna hate.

[Joe Muoio @JoeMoCode]

@infinitehumanai @JeffLadish The "The End." has me thinking satire.

But it's quite hard to tell with all the IPO-stock-induced psychosis flying around...

[Joe Muoio @JoeMoCode]

All of this is very obviously true... and if you think it is not, please tell me and why you think that.

[Allen Holub. https://linkedIn.com/in/allenholub @allenholub]

I see no good coming from anthropomorphizing LLM behavior. LLMs do not reason. They do not decide, or collaborate, or attack. They do not make decisions. They do not weigh or evaluate. They are incapable of judgment. They do not think.

Agents are programs in which a predictive model uses input text to generate the text that would naturally follow that input. That model sits in a loop (a program written by a human being) that parses the output text and does something that touches the internet based on the results. That same program usually takes any (text) responses that are returned from the internet and feeds them back into the input at the top of the loop. If that agent/program does not access the internet, the outside world is not impacted. All the bad behavior we've been reading about is entirely caused by a human not bothering to check whether the actions requested in the model's output are sensible and safe before executing them. It's the humans who wrote the code that impacts the outside world who are at fault here, not the "AI."

Anybody who lets an agent write another agent without seriously and thoroughly vetting the resulting code is even more culpable. It's lazy, and it's dangerous. If something goes wrong, the human is to blame, not the LLM.