Notices by Jason Gorman (jasongorman@mastodon.cloud), page 2

LLMs are a dead end. Not reliable enough for most use cases, and never likely to be much more reliable.

A human with expertise will always be required to instruct and monitor the work. And, as we're seeing, the productivity gains just seem to evaporate when they're actually measured because these models tend to create more downstream work than they save us.

The best summary of "prompt engineering" I've seen so far.

"We couldn't think of a way to add value, so we took value away and you'll have to pay extra to get it back."

We live in an age where "premium" means "Not deliberately shit". What a time to be alive.

I've just invented Vibe Managing, and I confidently predict that 6 months from now software engineers won't need managers anymore.

When I see posts on social media complaining about the reliability of LLMs in generating code, there are 2 distinct kinds of replies from the "A.I." boosters that amount to:

* "You must be throwing the dice wrong."

* "Are you using the latest dice?"

They are never accompanied by a demonstration of the "right way".

Comprehension Debt: the time it will take developers to understand code they didn't write when they have to change it or fix it.

Perhaps the solution to the Fermi Paradox is that technological civilisations use up all their planet's resources trying to make an intrinsically unreliable technology reliable enough to answer sales queries.

The solution to poverty is money.

The solution to hunger is food.

The solution to homelessness is homes.

The solution to illness is healthcare.

The solution to conflict is dialogue.

We don't need a fucking chatbot to solve these problems for us.

This is the best summary of "A.I." LinkedIn ever.

LLMs aren't linters, folks. They report imaginary problems, sometimes in imaginary code, and they miss *tonnes* of problems.

I'm not at all convinced that vibe-coding doesn't crap out after a few hours, TBH.

Hypothesis: we *know* that vibe-coding craps out at scale, and they can reach that scale within a few weeks at the latest.

Therefore, if the claims are true, and thousands of vibe-coded products are being launched every day, why isn't my phone ringing off the hook with desperate founders stuck out at sea on rapidly sinking homemade rafts?

I think they might not actually exist.

I think the people claiming they're vibe-coding real products are actually just touting for business as "AI experts"

Maybe should have asked ChatGPT about "the tortoise and the hare"

I don't need to say this often, but someone who does whatever the person signing the cheques asks isn't a professional. They're a mercenary.

The Fourth Law of Robotics is that there are two R's in "strawberry"

There seems to be a consensus that the performance of LLMs and "agentic" coding assistants deteriorates as the code base gets larger, to the point where they often become more hindrance than help.

Is there any data on this effect, though?

As a technical coach, I could be replaced by a t-shirt.

On the front it would say "YOU'RE PROBABLY OVERTHINKING THIS", and on the back, "NOW RUN YOUR TESTS"

@feld @jwcph Going by the hard evidence, nobody's had an LLM do that, either. Not without them fixing it.

What makes LLMs work isn't deep neural networks or attention mechanisms or vector databases or anything like that.

What makes LLMs work is our tendency to see faces on toast.