iiinova — where your digital self keeps evolving

What a weird and sad way for our industry to devour itself

010

I don’t know when all the LLM fans decided to start pretending hallucinations don’t happen all of the time? Was there a memo

000

Every day on here there’s some main character claiming AI is about to end however many various careers followed by another report like this

121

Anyway bottom line I think if people are taking the time to fully understand what the AI is doing it's probably not really saving any time, and if they're not doing that then they're letting digital rot accumulate in the walls. IMO the Meta findings I linked bear that out. Sorry for the ramble

100

None of this is a criticism of you, I hope that's clear. Mainly I'm bemoaning the deep, deep problems in the tech industry that have been with us a long time. In a lot of ways I think we probably deserve AI (derogatory)

100

I've watched multiple places succumb to that and go out of business because at some point it's too costly to start over and it's too costly to keep going. I have yet to see any attempt to address this re: AI. It seems like most people assume it will be immune to it or just don't consider it at all

100

So over time you build up this tech debt and never budget anything to address it and there's always something new the bosses or clients want and it never gets done. And you can go on like that for awhile! But eventually you get to where the codebase is like wading through mud to get anything done

100

It's not specific to AI, you can have like, a junior copying code they don't understand, or a gun for hire writing stuff nobody does. Which you can absolutely correct, just like you can correct AI code, but it takes time and that mostly negates the whole "we did this to go fast" thing so nobody does

100

Ah yeah I know that pain. Been there in the pre-AI days. I guess my experience is that shortcuts of this nature are almost never worth it in the long run

Would you be willing to share an example of this kind of work? Honest question

100

I’m not saying it can’t write useful code, but it can’t write useful code without either a) taking as much time as humans would take or b) causing equal or greater negative impacts, so what is the point bsky.app/profile/jeff...

140

The same people also regularly claim to use LLMs for things they cannot do, which makes me think it goes beyond bljnkered for some. Like Claude is not reviewing your code, man. You fell for Anthropic’s marketing

040

We’ve been asked to not argue here so reply on this thread if you’d like. But all of those things require either reasoning or a deterministic process, LLMs are neither bsky.app/profile/jeff...

Can you explain how an LLM is useful for coding

120

The only thing that would make LLMs useful is if they had human level reasoning. They do not and never will, new models cannot change that

People say exactly this, every six months. Just one more model please bro

1100

Tim Ellis, TDSB Ward 2 candidate 🍁 @djdynamic.ca · 17/09/2026

It hasn’t bsky.app/profile/jeff...

140

Reposted by Jeff Smith

If all you have is a hammer (a wildly inflated sense of your own intellectual prowess) everything looks like a nail (a logic puzzle that everyone else was simply too stupid to solve)

091

Jeff Smith @jeffsmith.bsky.social · 13/09/2026

Once again, “hallucination” happens at an insanely high rate and it is not solvable. How are we still having this conversation

051

And yet something about AI tools leads a percentage of its users to assign it abilities it does not have: it *feels* fast therefore it *is* fast. At the end of the day the value of AI as a coding tool is literally vibes

0150

Step back from the confusion around "AI" for a moment and this should all seem very obvious. We've known for decades that "more code faster" does not mean more value

1152

This is what we mean by "meaningful" acceleration; an increase in actual value that isn't negated by a loss of actual value (stability, security, etc).

1130

Meta reportedly abandoned an AI-focused restructuring plan that would have laid off thousands - Engadget

More recently, Meta found that while code changes "were up 220 percent year over year", "new or improved features [...] rose just 36 percent" while "technical and security incidents rose 40 percent" and "time employees spent on those problems grew by 70 percent." www.engadget.com/2244816/meta...

engadget.com

The problem with Meta's AI restructuring plan is the AI wasn't good enough.

1152

The claim is that using AI to code is beneficial because it makes development faster. But typing code is not the bottleneck, it never was. Making decisions is the hard part and something that LLMs are and always will be incapable of doing

1190

Now a lot of engineers I've talked to at this point will say "well of course! That's why you have to do the thinking/make the decisions!" But I find this to be a dishonest retreat

190

First: technically speaking, it does not "write code". It generates text. This is important because "writing code" implies that it can understand the code it's generating, that it can choose the proper architecture or approach based on the needs or goals. It can't do any of that.

1170

The rhetoric has shifted once again, this time from "it's a tool" to "this is so obviously right you're delusional or a liar if you don't agree" so I will once again attempt to explain why AI is not, in fact, good for code

1488

Probably nothing to worry about

000

It’s something about the way the narrative can adapt to any contradiction, the way god/AI can never fail, can only be failed. The same flavor of thought-terminating cliches

100

Jeff Smith @jeffsmith.bsky.social · 29/08/2026

Idk if I can quite articulate the how and why but as someone who grew up in a religious cult, talking to people still in it and talking to fans of AI feels exactly the same

100

Upstart numbers. Rude numbers

000

Jeff Smith @jeffsmith.bsky.social · 29/08/2026

We have a weirdo 370x208 resolution, but I’ve found it looks good at most sizes with a decent pixel shader!

Jeff Smith @jeffsmith.bsky.social · 28/08/2026

This one in particular has been getting to me. Watching people abandon best practices and curiosity to eagerly self-delude themselves is not great

000

Jeff Smith @jeffsmith.bsky.social · 21/08/2026

Everyone *feels* like it makes them faster but that is a perception not borne out by studies. LLMs cannot review code. They cannot write documentation. Can we all stop assigning them human level reasoning please, you are falling for Anthropic’s PR

000

Jeff Smith @jeffsmith.bsky.social · 21/08/2026

Hallucination means everything it generates has to be understood by a human, reviewed by a human, corrected by a human. Every decision at the most granular level has to be made by a human. It means there are no productivity gains at scale

120

Jeff Smith @jeffsmith.bsky.social · 21/08/2026

A thing making me quietly insane is how we’ve all for some reason decided to start pretending the hallucination problem has been solved when it very much has not

100

Reposted by Jeff Smith

Alexander King @literallyaking.com · 19/08/2026

I think maybe the reason this didn't happen is that promises like the former are quickly verifiable as something it can't actually do, whereas promising to replace all labor *in the future* is a goalpost that's conveniently mobile. (Since, largely, the real customers are VCs etc, not people)

1194

Jeff Smith @jeffsmith.bsky.social · 08/08/2026

I would push back on your last point. I know plenty of smart engineers who have no use for it; I think the idea that this particular tool is effective for this particular job is an assumption that’s yet to be proven. As Suresh correctly points out we still have no measurable evidence that it is

000

Am also curious how practice can help with a nondeterministic process

010

I remain skeptical that the juice is worth the squeeze. You still have to make every decision yourself, so where is the speed benefit? Code isn’t the problem for me, I can type very fast :)

I agree with most of this! It could be we’re more aligned than we think, and the difference is in the type of projects we’re working with

100

I’ve seen the Claude Code leak. I’m sorry but I simply do not believe you

050

AI Mania Is Eviscerating Global Decisionmaking

No insult taken. I read this today and found it lined up pretty 1:1 with what I’ve seen in the wild: hermit-tech.com/blog/ai-mani... I can’t speak to what’s going on at Apple or why Torvalds believes what he believes. All I can say is that making software is making decisions and LLMs can’t do that

hermit-tech.com

The reality of how much damage AI mania is doing to our ability to run institutions effectively.