Jerry Kindall draws conclusions about AI
Nico Neugebauer’s free Claude Code plugin, Heard, makes Claude read back what it thought you told it to do before it does anything, just like when the chef calls out an order in a professional kitchen.
FFO: Claude Code, The Bear, and the brigade system.
There’s no single ancestral model from which GPT, Gemini, Claude, and their rivals descend like digital offspring. But there is a mother architecture: the transformer. It comes from an eight-page paper published in 2017 by a team of Google researchers, with one of the best titles in the history of computer science:
“Attention Is All You Need.”
(more…)I don’t always work with AI, but when I do, I use Perplexity. For one thing, it gives me access to a number of popular models, including GPT, Claude, and Gemini… no need to have separate subscriptions to all those models. For another, the company offered a complimentary 12-month Pro subscription last December to all PayPal users, so I’m still in my free trial. Free beats cheap.
In Perplexity, you can fill in personal preferences to establish background and standing instructions. (Pretty much all AI chat interfaces let you do this.) All models use the same preferences, and it’s interesting to see the differences in how various models interpret your instructions. My instructions include this passage:
Prefer declarative sentences and active verbs. Prioritize clarity, precision, and logic. Avoid wordiness: fluff, padding, throat-clearing. Don’t flatter me or try to sell me. Make every word earn its place, each thought follow from the last. Be certain of facts. Make each sentence feel necessary.
There is one model that follows these instructions better than the rest, to the point of being borderline brusque at times, and that model is Grok. As problematic as Grok is, it consistently challenges my writing and ideas in the way I want them to be challenged. It willingly says “that’s a bad idea” or “don’t hedge.” It contradicts me without hesitation.
In short, it’s a good foil to the tendencies that Claude (and to a somewhat lesser extent, GPT) encourage. Even after being admonished against blowing smoke up my ass, Claude often tells me my ideas are the Best. Ideas. Ever. So my current AI practice is to brainstorm with Claude, test drafts against Grok, and do final review with GPT.
How do you do it these days?
I use Perplexity for working on my blog because it gives me access to most major models. Perplexity’s Best mode automatically chooses a model for each of your queries based on its opinion of which model would give the highest quality response for that particular prompt. In Best mode, then, the model you’re actually talking to can change from prompt to prompt.
Because large language models are effectively stateless, models consider the whole conversation, from the very beginning, in order to respond to your prompt.1 In Best mode, whatever model Perplexity chooses will get the entire conversation history, including previous responses written by a different model.
I thought that perhaps an LLM might balk somehow at a transcript that claims a model gave a response that in reality it would never give. So I asked Perplexity about that, naturally. Turns out that’s not a problem; LLMs take the conversation transcript at face value and do what they always do: use it to predict the next response. In fact, you could synthesize a transcript with complete non-sequiturs for the model’s previous “responses” and it’d work perfectly well, in the sense that you’d get a response that’s reasonable for the context.
One implication of this modus operandi is that if a conversation shows the model responding one way in past turns, it’s more likely to respond the same way in the future. (You could use this to send a few fake conversational turns that reinforce instructions in the system prompt.)
And that brings us to why Perplexity keeps writing in Chinese. In Best mode, Perplexity can use DeepSeek to respond to some requests, even though you can’t select that model from the UI. And DeepSeek, being a Chinese model trained on lots of Chinese text, has a penchant for randomly responding in Chinese even if you’re prompting it in English. Once that pattern has been established in a conversation, Perplexity tends to keep doing it. Even other models do it once DeepSeek has done it a few times!
I often use Perplexity for learning new things, because it can search the Web and synthesize a concise overview of the basics of a topic. Then I can probe for more detailed information about some aspect, or present the AI with my understanding and have the machine poke holes in it.
That’s how I ended up in a conversation about cologne with Perplexity. I’ve recently become interested in the topic, and I’ve used Perplexity to explore styles and notes and to get recommendations of specific fragrances to try. In this conversation, I happened to mention that Valentino Born in Roma Uomo seemed a heck of a mouthful for a fragrance brand, especially once you tack on brand extensions (e.g. Valentino Born in Roma Uomo Purple Melancholia).
Perplexity pointed out correctly that “uomo” is Italian for “man” (identifying the male-coded version of the fragrance, as opposed to the female-coded “donna” version).Then it hit me with the revelation that the Valentino brand came from the founder of the house, Walter Valentino.
Wait, what? Walter?
A quick Google revealed that there is in fact a fashion designer named Walter Valentino, so I shrugged and assumed the matter was closed.
Me: I thought you must have hallucinated Walter, but it’s true! I Googled it!
Perplexity: Okay, I was 100% wrong — my apologies! Valentino Garavani’s full name is Valentino Clemente Ludovico Garavani. His first name is Valentino, not Walter. […] Walter is just… I don’t know where I got that from. Maybe I misread something or it was a complete hallucination. There’s no “Walter” in his name at all.
Perplexity then gave me some background on the real founder of the Valentino fragrance brand, including the fact that he died earlier this year. All accurate.
But wait!
First actual case of a human being found hallucinating like an AI?
Of course, I ran a draft of this article by Perplexity, and it pointed out that both of us “made identical epistemological errors.” Turns out confirmation bias doesn’t care whether you’re made of neurons or matrix multiplications.
J. Michael Straczynski, a man who knows a thing or two about storytelling,1 has a two-part essay on his substack about why AI has not yet produced a great manuscript, and might never.
Silence where a story might have been (part 1)
Silence where a story would have been (part 2)
AI is already being used to write manuscripts in collaboration with human writers. We know this because a handful of writers have been caught at it. But almost certainly there are writers who weren’t caught, and whose AI collaborations have been published.
He’s right that AI won’t write a story on its own. “Write me a story” is a laughably vague prompt. You’re leaving the details up to the model, which will produce an average of all the literature it knows about, and that’s statistically very likely to be mediocre.
If you prompt better, you’ll get better results. If you iterate, providing yet more specific instructions, you’ll get better results still. More on that is coming soon (I’ve been experimenting).
But you’re still the one steering the story machine. Straczynski’s point is that AI doesn’t do anything without a human telling it to. By its very architecture, it will never observe a weird little moment of human interaction and think, “I should write a story about that.”
I think it’s way too early to say conclusively that a machine will never be able to have that experience. But such an AI will be a lot more advanced than anything we have now… and it’ll probably have a body.
In 1993, a company called MAI Systems sued Peak Computer, a computer repair firm, for copyright infringement. Peak’s technicians had turned on MAI clients’ computers to diagnose them, which loaded MAI’s proprietary operating system into RAM. MAI argued that loading its software into memory—even briefly, even just to make sure the computer was working, even though there was no way to start the computer without loading the software—constituted an unauthorized copy. Because the Peak technician was not licensed to use that software, MAI said, Peak had committed copyright infringement.
They won.
(more…)Michael Andrews has an interesting post about some of the credibility issues surrounding generative AI: phantom authorship, potential boobytraps, and countermeasures against them. A key insight is that most LLM knowledge isn’t from primary sources of facts, but rather from secondary sources’ simplified explanations of the primary sources. The latter are simply more numerous. So sometimes, a LLM’s knowledge lacks the depth you’d want.
If you’ve ever watched a television news report about something you’re an expert in, you already know how wary you should be of secondary sources.
The aphorism “quantity has a quality all its own” is usually attributed to Josef Stalin, though the earliest traceable source appears to be an American defense newsletter from 1979.
Whoever said it, they were talking about armies. More soldiers don’t just let you fight the same battles better. Past a certain size, your army can fight a different kind of battle entirely, with new strategies and tactics. The observation applies at other scales, too: firing a machine gun is a very different experience from firing a musket—especially if you’re on the receiving end.
The aphorism might as well be the thesis statement for one of the most surprising discoveries in modern AI research. I touched on it in my last post when I compared Mark V. Shaney’s capabilities to modern LLMs. The discovery is this: bigger language models don’t just do the same things, only better. When they grow large enough, they start to be able to do entirely new things.
(more…)