This is like measuring the increasing speeds of cars in the early years and extrapolating that they would be supersonic by now by ignoring the exponential impact that air resistance has.

"performance"

Then why share it?

Is the performance increase related to computing power? I suspect the undelying massive datacenters running the cloud based LLMs are expanding at a similar rate...

2 X 0 = 0

This is such bullshit. Models have already consumed all available data and have nothing left to consume, whole needing exponentially more data for progressive advancements

All that power used for a fucking coin flip.

So we can mock it!

How is completely fucking up literally 50% of the time outperforming exactly???

Do you not see any value in engaging with views you don't personally agree with? I don't think agreeing with it is a good barometer for whether it's post-worthy

That sounds like a coin flip, but 50% reliability can be really useful.

If a model has 50% chance of completing a task that would cost me an hour - and I can easily check it was completed correctly - on average, I'm saving half of the time it would take to complete this.

That said, exponentials don't exist in the real world, we're just seeing the middle of a sigmoid curve, which will soon yield diminishing returns.

You see, in 7 months, they'll fuck up literally 100% of the time! Progress.

Apparently, throwing more data at it will not help much from now on... But anyway what they're saying, I can't trust the snake oil seller, he is suspicious...

Is it just me, or is this graph (first graph in the article) completely unintelligible?

The X-axis being time is self-explanatory, but the Y-axis is somehow exponential time but then also mapping random milestones of performance, meaning those milestones are hard-linked to that time-based Y-axis? What?

Very good analogy. They're also ignoring that getting faster and faster at reaching a 50% success rate (a totally unacceptable success rate for meaningful tasks) doesn't imply ever achieving consistently acceptable success.

Air resistance has cubic not exponential impact

Good point, thank you, I figured that sharing poor scientific articles essentially equals spreading misinformation (which I think is a fair point either), but I like your perspective either

Then why do I feel like it's programming abilites are getting worse? I've stopped paying for it now because it causes more frustration than anything else. Works for simple "how can I simplyfi this code" queries when my head hurts, but that's about it.

So only 10 years until it isn't a ressource hog anymore...

I guess the value is that at some point you'll probably hear the core claim - "AI is improving exponentially" - regurgitated by someone making a bad argument, and knowing the original source and context can be very helpful to countering that disinformation.