Alexander Schmidt-Lebuhn

I am so proud that I understand that reference.

It is well possible that some different, future iteration of AI will be reliable and able to replace human labour in more areas than slop art or spam generation. But we just don't know if such an AI will be affordable either. Human brains are really astonishingly energy-efficient "per token".

This Is How Much Anthropic and Cursor Spend On Amazon Web Services So, I originally planned for this to be on my premium newsletter, but decided it was better to publish on my free one so that you could all enjoy it. If you liked it, please consider subscribing to su...

Second, from what I am reading, LLMs are extremely expensive to train and run. The providers are setting billions on fire every year. It isn't clear that users will ever be ready to pay as much as it would cost to make this profitable.
www.wheresyoured.at/costs/

Based on a recent discussion about 'AI' with other scientists, here are the two things I'd really like everybody to understand.

First, from what I am reading, 'hallucinations' are part of how LLMs work, so they are never going away, so you cannot use them for anything that requires reliability.

This is also just not reflective of how people do things. In my discipline, yes, the first author is generally the postdoc or grad student who led the study, and the last author is generally the supervisor or whoever got the funding. But there are other fields that simply do alphabetical order.

I remember one of the last two times I decided to give Windows a chance. Bought computer, found it needed 45 min to install updates, so I did dishes. Nearly had a heart attack when Cortana suddenly said something, bc I thought some stranger was in the room. Wiped the machine and installed Ubuntu.

It is tens of thousands of years in the future, and there are still the same 'races' as today, gender roles haven't changed, and people still consume the same drinks. Flying between stars with a spaceship is like driving a caravan between cities, only you have an atomic engine instead of petrol.

This thread resonates with me a lot. When I read the Foundation novels, I kept wondering where that great scifi writer was that everybody sees in Asimov. Even if you leave aside his misogyny (which you shouldn't), he had very little imagination.

This would be obvious supervillain stuff if the robots worked out.

Now to see if they will work exactly as well as the boring tunnels, full self driving, Mars colonies, and Grok becoming a superintelligence.

Bioinformaticians and their software names... surely there will be no confusion between this and RogueNaRok 😀

I for one and pretty sick of the “I voted for the Trump in my imagination” excuse.

Voters are never going to study policy statements on websites; that’s not a realistic standard. But if you can’t even get the vague *gist* of what a candidate wants to do, that’s a You problem. Skill issue.

May be time for everybody to internalise that most journalists are in one of three lines of business: generating online engagement, including rage clicks; telling the reader what they want to hear; or telling the media owner what they want to hear. And I get it, people have to pay their bills.

The average age of U.S. homebuyers is now 56, up from 49 last year.

In 1981, the year trickledown economics began, it was 31.

Ooof. Just read that thread. I guess my attitude is that if I'm writing a grant on a topic, I should know enough about it that I don't need AI to find citations and develop a question for me. If I don't know what a specific gene/pathway does in a process, should I really be writing a grant on it?

The usual thought exercises, just for fun:

If you had predicted in 2023 that Trump would demolish part of the White House, how would Sensible Centrists in the USA have treated the prediction?

If Obama had done this, how would US conservatives have reacted?

It is fair use if somebody cites my work with attribution. It isn't fair use if they use my work to train an AI with the explicit intention of making a profit from making me unemployed. Not a lawyer, just a moral argument.

Apart from the incompetence and laziness of the professor, what is so bizarre about this is the idea of some random professor having their own textbook that they teach from. In my field we had The Textbook that was co-written by several top people in the field and was used across many universities.

Sadly, I know! I saw a pre-print fairly soon after ChatGPT came out that discussed the reduction in activity on SO.

What I am saying is, will they continue using ChatGPT or Claude if subscription is $2000 per year? Locally run LLMs may work, but not for unskilled vibe coders outside of corporations

Low-complexity, one-off scripts is the niche for this stuff, yes. Now one will have to see who is willing to pay what level of fees for a few one-off scripts per year when they could also google their problem and click onto the Stackoverflow page that comes up.

I sympathise with this perspective, but I am not sure how well generative AI would be able to work as a tool with some uses if it were restricted to training only on consensually acquired training data. How much would be left?

Even as a teen in late 80s/early 90s I noticed the implications of making the heroes disruptive entrepreneurs who chafe against rules and the main villain an environmental protections officer. That decade was when neoliberalism took over but also when the green parties got into parliaments, so...

Unless when Andreesen says "filmmaker" and "idea", he means this, of course.

bsky.app/profile/beau...

Even if generative AI could do this, it would still be silly to believe that all that matters is the idea. If the "filmmaker" in this scenario had no understanding of writing, direction, editing, etc, they would fail spectacularly regardless of how well the AI slave implements their instructions.

Question folks should ponder: what happens when no one contributes to StackExchange anymore because no one goes there anymore and the training data for the AI runs out and gets stale?

Not even a professional coder, so my problems are simple. On my first try, the script it produced didn't do the thing I requested. On second, I did four iterations of explaining to it what its previous attempt did wrong before I got the intended result. Save some time, maybe. Transformative? No.

This internal contradiction is what gets me about his argument. Either you have to invest enormous time and care into using it well, but then it isn't all that transformative to most people. Or it is transformative, but then it should be easy to get impressive and reliable results. Which is it?

Also noting that "you must not have learned how to use it effectively" both is of the same shape as "[failed idea] has never *really* been tried" and stands in tension with the claim of transformativeness. If it is that difficult to figure out how to use LLMs, how useful can they be to most people?

I tried coding, text extraction into table, literature summary, and more. LLMs either faceplant spectacularly or are too unreproducible.

Some colleagues find LLMs useful to document their code, spot typos, or create meeting summaries. And there's slop art/text.

Is that "transformational"? Hardly.

Huh. I have never even thought about how Star Trek would have come across if seven bridge crew members had all been trying to talk to the ship's computer at the same time.

Reductionism as the author defines it and emergence are both true and not alternatives as he claims. However, reductionism is false at a *practical* level, because there is no realistically feasible way of explaining the existence of cross-stitching entirely from the four fundamental forces.