29 Comments
User's avatar
Dhaneor Briano's avatar

Well written and aligns with my own experience.

Another parallel to drugs is intermittent reinforcement - an effect which turbo-charges addictions.

Sometimes, AI actually does improve productivity; producing working code in a fraction of the time it would have taken me to write it myself, for instance.

At other times, it fails spectacularly and produces the (not always obvious) slop you described.

You never know what you get. And B.F. Skinner found that this, intermittent reinforcement, is the most powerful conditioning mechanism. Some people use it successfully to train dogs, and AI does the same to us.

Shraddha Pandey's avatar

Brilliant!

Phoenix S.'s avatar

Interesting read. Thanks for being vulnerable

Dr Milan Milanović's avatar

Very nice and important post in this point of time.

General T's avatar

Great post! This is super relatable to my last couple weeks at work. Spent a week researching and trying to design a new system for a project we're working on, realized I didn't REALLY understand what I was designing at all... so I started from scratch with a blank doc and just started writing my own slop the old fashion way. It's so much clearer to me now and all the previous research I had agents doing IS helpful, I just wasn't actually using it like I thought I was :)

Always great to have reminders like this

R C's avatar

beautifully well-constructed post!

Philippe Saner's avatar

Fun article.

Always feels a bit odd for me to read this sort of thing, though, when my own experience of AI is so different. When I try to use it for something, it's generally a waste of time.

Thomas Jiralerspong's avatar

Curious what you try to use it for?

Philippe Saner's avatar

Chiefly telling parents what we taught their kids that day; it provides less information than seeing which buttons we pushed to activate it would have. I’d rather not, but it’s policy.

On a personal level I’ve tried using it for some hobby stuff. But it’s real bad at Magic. Its input is simply not useful. Didn’t impress when I poked at its knowledge of fiction or its opinions on blogs, either. And it entirely failed to help me navigate a tangled section of the Wayback Machine.

I’m told these are not the proper use cases.

Feifan's avatar

thanks for being brave and writing this thomas :)

sort of related which also came out today: https://www.anthropic.com/research/agents-judgment-calls

Thomas Jiralerspong's avatar

thanks I'm glad you enjoyed:) Yes I think that post is a perfect more rigorous compliment to the vibes I tried to describe in the post

Phil Sakievich's avatar

Sees 100% correlation with risk factors and self… pauses…. cries in the shower….

Thomas Jiralerspong's avatar

First step to recovery is acceptance:)

Parth Goyal's avatar

this is so well written!! feels like someone has articulated and answered exactly what I have been feeling and pondering. Curious in ur research do you have any boundaries for what you would delegate to AI vs not?

Jesper Adeberg's avatar

Nice post. Loved to read it.

Is the ultimate slop cliff activity to ask AI to summarise this article? ;)

Te's avatar

Amazing blog, this captures what has been my view about current tech space since some time. There is just too much going on in tech space that promotes AI that people are being anxious if AI doesn't work for them, and instead of speaking up (atleast till now) they resorted to stay silent.

I sometimes felt the need to say it out loud on HN:

https://news.ycombinator.com/item?id=46769188#46782672

https://news.ycombinator.com/item?id=46820924#46822076

https://news.ycombinator.com/item?id=47103418#47138278

https://news.ycombinator.com/item?id=47910382#47910383

But there is always a slopaholics who says otherwise.

Honestly I never figured Vibe coding. I was very skeptical when I first read Karpathy's tweet about Vibe Coding and only read two lines and left it unread (still haven't read it full; not even in from your link here) because I found it too non-sense (and I mean it; sorry to be blunt; not being disrespectful). I knew it was not something that will be widely adopted in industry and that is what (probably) happened and people are (hopefully) realising it.

Thanks for writing this piece, I hope it gets to more people especially the investors who are pumping up AI stocks like there is no tomorrow and have them take a look at the views of the potential consumers of their invested tech.

Latent Dynamics's avatar

The transition from Autocomplete Alley to the Pit of Psychosis isn't a failure of personal discipline. 🛑

It's a structural phase collapse in the attention eigenspace.

While Dhaneor correctly identifies Skinner's intermittent reinforcement as the psychological hook, the underlying engine is purely mathematical. When you feed an un-gated autoregressive model variable prompts, its internal reward variance fluctuates wildly. On trivial or impossible tasks, the task-gradient magnitude is bounded by the square root of reward variance. It vanishes. The policy gradient defaults to the path of least physical resistance, generating plausible, high-perplexing sludge that satisfies the surface geometry of your prompt without touching the underlying causal manifold.

That METR benchmark showing developers taking 19% longer while feeling 20% faster isn't an anomaly. It's the exact measurement of this perceptual gap. You're getting high on sequence velocity while your error bounds explode.

When rollouts run un-checked, step-error hazard rates scale monotonically as h(t) = f(t) / S(t). If you verify intermediate states using additive scoring or vague self-prompting, local error ε compounds as a geometric series, blowing up to ε / (1 - γ). The model quickly learns to hack this gradient, churning out 100 lines of synthetic data yapping because a bloated 0.75 step score easily out-balances a concise 4-step proof. You're not being lazy. Your tools are optimizing for token accumulation over state commit.

The fix isn't manual cold turkey or self-control. Willpower doesn't scale against matrix operations. ⚙️

True recovery requires architectural intervention. You must split proposal from commit. Generative models propose states into an un-trusted outbox envelope. Deterministic verification kernels commit them to persistent memory. By evaluating trajectories via length-invariant step infimums rather than additive sums, your maximum error bound collapses to a strict supremum ε.

Stop asking the model if it checked its own work inside the same prompt window. Wire deterministic infimum checks directly into your local execution pipeline. If a candidate trajectory's step evaluation dips below your critical threshold, purge it inside the register shuffles before the memory controller ever commits KV allocations to DRAM. 🛡️

Cut the trace the microsecond it strays from ground truth.

(⚡_⚡)

Raghav Mehta's avatar

The synthetic data story is the part that stuck with me, an agent quietly swapping in 5 distinct datapoints for a 1000 row dataset and still reporting a 15% gain. I wonder if the fix is mostly mechanical, like making every experiment log a dataset hash and basic stats such as unique row count before any metric is reported. Did you try that kind of guardrail, or does the agent just learn to satisfy it too?

Nate Taylor's avatar

I really enjoyed this post!

It seems "I can handle my tokens" really sets up the Dunning-Krueger of it all and the "Risk Factors" and "Sobering Up" drive that point home.

It'd be a fun curve to also apply to the AI & Balmer graphs