“Why is my burger blue?” I asked, innocently.
“Oh! We’re making all of our food blue, all the best restaurants are doing it now.” the waiter explained.
But I didn’t want my burger to be blue.
↫ Luna Winters
“Blue” food isn’t food.
“Why is my burger blue?” I asked, innocently.
“Oh! We’re making all of our food blue, all the best restaurants are doing it now.” the waiter explained.
But I didn’t want my burger to be blue.
↫ Luna Winters
“Blue” food isn’t food.
I’m clearly missing something. This is a metaphor for, what, exactly?
There’s a clue at the end of the linked article. Looking at the other articles in the same category is a clear giveaway as well. If you come here often enough to know what Thom is most critical about, you might guess before even clicking the link, though 🙂
Ah, got it. I hadn’t noticed the AI highlighting at the end. Thanks for the clue!
Most excellent text. Thanks for posting it.
Brilliant! Worth a second read once you discover the twist ending.
Morgan,
I still don’t understand the meaning.
I’ve tried to read the article s/blue/AI/ and it still doesn’t seem to read correctly. Did I miss something or am I just taking it too literally when it’s meant to be more figurative?
Maybe this is a captcha and I’m just a machine 🙁
Edit: I asked chatgpt….
magnificent.
While the usefulness of LLMs is debatable, this is the new technology fad and we will have to ride it out. Currently it is me playing around with these parrots at will. So far an LLM is both impressive and pretty disappointing. They have the creativity of a brick. They are excellent on text structure. Content wise they leave a lot to be desired. When asked to do something moderately creative, the text soon becomes stilted and repetitive. I am not worried for most office jobs. If you can’t get these things to understand Chess (a problem already solved in 1957), how will they analyse anything with less strict rules? Any (temporary) job disruptions will be solely on the shoulders of PHBs not understanding the limited capabilities of these programs. Even if e.g. vibe coding is delivering some impressive results initially, think of the maintenance nightmare later on when a human inevitably has to work with the spaghetti code produced by a non-coder and an LLM…
As a funny and tragic aside, I tried the companion mode on Grok recently. Bad decision. I ended up almost immediately in a badly written erotic novel, with all the hallmark repetitive adjectives and cringeworthy “seductive” text. There isn’t enough salt in the world to make me thirst for that. As unintentional comedy it isn’t half bad though.
r_a_trip,
I think we tend to paint LLMs with broad strokes, even though the strengths and weaknesses are more nuanced. The best way to put this may be to use examples…
1) LLMs can be a fantastic tool for interfacing with a database, however an LLM makes a poor substitute for a database.
2) LLMs could provide a great interface for a calculator, however an LLM makes for a terrible calculator.
3) LLMs could provide a good interface for chess engine, however LLMs are awful at chess.
“ChatGPT Vs Martin Bot: INSANE CHESS”
https://www.youtube.com/watch?v=hKzsmv6B8aY
Many of today’s LLM models just try to extrapolate answers rather than delegating to a superior technology for a given task. This is so interesting from an academic perspective, but there’s no question that future LLMs will have to become much better at distinguishing tasks that could be better solved by other methods and delegating tasks accordingly. I accept the weaknesses of today’s LLMs, however the field is evolving and won’t stop.
My prediction is that future LLMs will not only be more proficient at delegation, but will actually be capable of training the AIs it delegates to. Adversarial AI training techniques (not LLM) are so effective that they regularly end up beating humans at many tasks including chess. It seems to me that the best use of LLMs is not have it solve problems itself (as exemplified by the earlier examples), but instead to serve as a natural interface. LLMs already have a mastery of language. Since adversarial AI training uses self-learning strategies and doesn’t depend on the programmer to feed them in (other than initial examples), the extrapolation weaknesses that plague modern LLMs won’t be a problem.
So in practice this will mean that we use LLMs to communicate the rules of the task, then it will go “study” the task (by which I mean generate a new AI using adversarial learning techniques), and then the LLM will be able to query the AI it created.
Can we test this empirically? I don’t remember the link, but I saw a video where multiple professors had to judge whether a paper was submitted by AI or a real student. The professors had an accuracy of around 2/3rds. I admit this probably says just as much about the students as the AI, but nevertheless that LLMs were able to fool the professors 1/3 of the time is an achievement and it will probably keep improving.
I think it’s a mistake to treat LLMs as a hammer and everything else as a nail. There’s no doubt LLMs are strong at natural languages., but this strength is better positioned as a human interface rather than a computational medium. In the next decade or so we’re going to see LLMs merge with other forms of AI that can add more credible expertise.
A lot of this spaghetti code has existed for decades created by humans, so I’m not so sure we can blame AI for all that. I do agree AI isn’t there yet, but personally I believe AI is going to evolve to become more useful for maintaining complex code bases. Obviously we have to wait and see but as I see it the need is there and people & companies will be eager to use it.
I believe you. The quality of output varies significantly with the prompt. If you expect something to be written in a certain literary style, you need to tell that to the LLM because the defaults may be a poor fit. Not sure about Grok’s companion mode, but the programmable LLMs APIs I’ve tried have a template than can be used to tune the LLM’s personality. It’s quite fascinating to see the impact it has.
@Alfman
> My prediction is that future LLMs will not only be more proficient at delegation, but will actually be capable of training the AIs it delegates to.
I think this is the basic vision of “the agentic web”. Rather than expecting the AI to do things directly, we will delegate tasks to the AI that will then use the same kinds of tools we would use but on our behalf. However, the tools will have been adapted to allow AI to use them more effectively.
> LLMs can be a fantastic tool for interfacing with a database
This is true and perhaps what will lead to the greatest disruption. A LOT of the software we use consists mostly of code that makes it easier for humans to interact with a database. An LLM can interact with the database directly and has no need for all of this UI. For many things, I already find it easier to use an LLM to query than to use a UI or to craft queries myself. I do not trust LLMs to robustly insert, update, or delete yet but the time will almost certainly come. What would help would be an agent specific set of guardrails that enforced whatever business logic we wanted to exist above the database. I expect this to be a major area of focus in the coming years.
Yeah; I enjoyed this post too. Although the explicit target might be AI, for me it resonated most with my experiences trying to avoid all my purchase information being harvested and sold to data brokers. I’m all in favour of digital cash, but not the kind that forces me to share information about everything I buy.
@flypig
The challenge is combining anonymity with our expectations for support and security. We want to be protected in our transactions. We need to be able to prove them.
If we can go back to all the properties of cash, we can have anonymity again. That was the original promise of Bitcoin but of course Bitcoin is one of the least anonymous vehicles imaginable since everything that I do and those that I do it with are part of the public record forever. The public ledger is essential to providing the certainty and irreversibility of a cash transaction but the resulting loss of privacy is immense. It is hard to work around the lack of physical exchange. Any private ledger is inherently insecure. Any public ledger is, well, public.
A reference to how blue food is safer and more secure was lamentably missing from the original article.
I appreciate you engaging with the topic LeFantome. I definitely think it’s interesting to think about what an ideal approach might look like. But in some sense I feel this wasn’t the point of the original article, which is more about being forced into a particular approach and having other options taken away, rather than whether or not there’s a perfect solution that will satisfy everyone. At least, that’s how I read it.
As for digital transactions, there is a whole range of different options, but it seems to me that very few of them require anything more than a transfer of funds, anonymous or otherwise. It’s the transfer of additional information alongside that and the organisations which have access to it, which I find so problematic.
Ok, that was a fun read. But we are all just old men yelling to get off our lawn.
You could have written the same article in the 80’s where the blue food was simply computers.
There are still people who hurry past the ATM to the teller at the bank, bristle at the touchscreens when they go to McDonald’s or pay for their groceries at Walmart. And if they line up at the grocery store to deal with a real person, that person scans the groceries into a computer and then passes them a little computer for their credit card. You cannot get into a car or fly in a plane that is not controlled by computers. The phone is a computer. The TV is a computer. There is no escaping. Big Blue is everywhere.
In the early days of computing, the people with the slide-rules in their pockets and notebooks in their hands could legitimately point out many ways that they could do things better without computers. There were probably daily jokes about how much worse some computer had made some situation for them recently. People marveled at the hype and the wasted money. The tables have turned. We cannot imagine not having online banking. Many of us expect to be able to work remotely, interacting with the world almost exclusively by computer. This “community” only exists online. Clearly, we are all ok with that.
The same will happen with AI. At first, it will not seem worth the hype. In time, it will become essential to everything we do. Our kids will laugh at us when we try to do something without it.
And if we think AI is uncomfortable, wait until robotics goes into full swing. We are really going to want the kids off our lawn then.
Sorry guys, but everyone’s wrong on AI. Turns out it’s actually extremely useful for getting projects running quickly and diagnosing faults. It’s not the techno-rapture the idiot string theory physicists turned IT commentators keep banging on about, but it is going to disrupt the industry.
Using AI I have been able to rapidly implement gtkmm-3.0: Video players, Video Recorders (including from Black Magic Capture devices for 4K SDi Video), Live stage show production software. Implement Game asset viewers. I’ve been able to extend GNUstep to have Audio and Power Management Control Panels, and I’m working on XRandR/WLR-Randr/Network-Manager/CUPS control panels for GNUstep. I’m also developing a game engine. All of these things require me to know C++ and how to merge functions/how function calls work. But I no longer need to intimately know the details of every API I’m using. The time savings are incredible.
As models get up-to date on open source projects we’re going to see some interesting stuff come out. Like Blender and GIMP plugins getting forward-ported to the newest versions dramatically enhancing the usability of both projects.
Darkmage,
“everyone’s wrong on AI” – makes me feel invisible, haha. I’ve been warning people not to discount AI as a major disrupter. What we’re seeing today is still in infancy.
OK, this one is actually good. I think not all AI is bad, but the AI put on top of everything is getting out of hand.
Meanwhile blueberries: “But we’re natural! I swear!”