Someone "actually working with it" checking in, if that matters at all to this conversation. I'm very bearish on the industry even if I think the tech is going to stick around.

If we separate the tech from the industry, it's clear one has some value (albeit very hard to say just how much) and the other is a lot of smoke and mirrors. This is not a healthy space.

One might point out that this is the story of AI since at least the 1930's. Impressive technology demo's ... wild investment, crash, bankruptcy ... but the tech remains and in fact has useful applications all around.

AI Winters. I finished school in 2008 and have seen it happen twice.

Convnets. LSTM (and various RNNs).

Both are in wide use today.

The difference is that every single one of those previous AI winters didn’t show up in the stock market valuations — not even a little bit.

The most spectacular previous issue was "Flanders Language Valley", I believe, but that was before 2008.

They had AI in the 1930s?

Vocoders, (voice encoder-decoder, they effectively compress voice) which were going to enable computers (the analog kind) to respond to human speech in callcenters (because editing/generating the compressed stream is hard, but actually doable)

Of course it was a joke.

But ... it was used. What they were used for a little bit is voice encryption, in military communication. Vocoder -> encrypt -> transmit -> decrypt -> Vocoder and you can "talk", if you don't mind the extreme voice distortion.

Oh and the talking clock on the telephone network. That, they got working.

> who are actually working with it only become more bullish

I have a feeling the word "actually" is doing a lot of work with this. I shipped AI facing user products a few years ago, then worked in more research focused AI work for awhile (spending a lot of time working with internals of these models). Then seeing where this was all headed (hype was more important than real work) decided to go back to good ol' statistical modeling.

Needless to say, while I think AI is absolutely useful, I'm bearish on the industry because current promises and expectations are completely out of touch with reality.

But I have a feeling because I'm not currently deploying a fleet of what people are calling "agents" (real agents are still quite cool imho), you would describe me as not "actually" using AI.

This really doesn't make sense to me. I see no world where Ai is so useful that the common man is willing to pay 100+ a month for it, but it's also a world where the common man has a job. There's too many people for everyone to have some niche job the Ai can't do.

And if such a job would carry a work week of, say, 5 or 10 hours?

Then they would be paid for 5-10 hours and have to ask the government for benefits.

In what world would a corporation pay a full yearly salary for 1/8th to 1/4 the labor hours? The current world already looks to labor as the juiciest place to cut cost for the profit margin.

That's not really how it worked out so far, the productivity is simply pocketed by the elite and never translates to shorter work week, salary increase or earlier retirement

https://files.epi.org/charts/img/235212-28502-body.png

Someone is working 20 hours a month and paying for a 100$ subscription on top of bills? And this isn't a isolated case, this is the expectation for the normal person? Is the job supposed to be real or is the government just giving out universal basic income while being petulant about people not working at all

That's not enough time to maintain skill. Experience would build very slowly in people working so intermittently.

With OpenAI, the most bullish analysts calculate that 40% of the world's population will be users in three years time, and they would still lose money on every sale. That's a bold bet.

would you bet $10,000 that the super majority (80% or greater) of current users of free APIs will be using a paid (20-200) one per month? if so let's set something up. we can set the time limit at January 1st, 2028.

Can I ask what you do? I suspect there is a type of job that AI excels at, and it makes everyone in that job unreasonably bullish on AI.

Well the problem is that even at $200 a month they're bleeding MONEY. FYI for every well intentioned skilled engineer using llms you have 100 lazy code monkeys shipping mountains of tech debt faster than ever before, 1000 people generating bullshit emails/tickets that could have been summed up in 5 bullet points, 1000 people role-playing with virtual friends/partners, &c.

You're in your own little bubble and completely oblivious to the thousands of man hour wasted every day to unfuck AI slop

And as it turns out 80% of users aren't willing to pay a cent and will abandon ship as soon as there is an alternative

Imagine throwing orders of magnitude more of compute at things - we may have things like a monte carlo tree search for LLM outputs using an LLMJudge that prunes the tree.

+ we can continuously let a LLM monitor our log files and alert/propose/fix issues 24/7. If intelligence becomes cheap enough this would be an enormous market.

Having a LLM run as "fact checker" /coach for everything that you write also would be a great addition.